persist node ids in .cache

brings back EXO_CACHE_HOME as always ~/.cache/exo/, and store the node
id in there. no random copies now!

.mlx_typings/mlx/nn/layers/convolution.pyi

@@ -32,6 +32,7 @@ class Conv1d(Module):
     """
 
     weight: mx.array
+    bias: mx.array | None
     groups: int
     def __init__(
         self,

.mlx_typings/mlx/nn/layers/linear.pyi

@@ -40,6 +40,10 @@ class Linear(Module):
         bias (bool, optional): If set to ``False`` then the layer will
           not use a bias. Default is ``True``.
     """
+
+    weight: mx.array
+    bias: mx.array | None
+
     def __init__(self, input_dims: int, output_dims: int, bias: bool = ...) -> None: ...
     def __call__(self, x: mx.array) -> mx.array: ...
     def to_quantized(

.mlx_typings/mlx/nn/layers/normalization.pyi

@@ -88,6 +88,9 @@ class RMSNorm(Module):
         dims (int): The feature dimension of the input to normalize over
         eps (float): A small additive constant for numerical stability
     """
+
+    weight: mx.array
+
     def __init__(self, dims: int, eps: float = ...) -> None: ...
     def __call__(self, x) -> mx.array: ...
 

.mlx_typings/mlx_lm/generate.pyi

@@ -73,9 +73,11 @@ class GenerationResponse:
     finish_reason: Optional[str] = ...
 
 def maybe_quantize_kv_cache(
-    prompt_cache, quantized_kv_start, kv_group_size, kv_bits
-):  # -> None:
-    ...
+    prompt_cache: Any,
+    quantized_kv_start: int | None,
+    kv_group_size: int | None,
+    kv_bits: int | None,
+) -> None: ...
 def generate_step(
     prompt: mx.array,
     model: nn.Module,
@@ -252,7 +254,14 @@ class BatchResponse:
 
     texts: List[str]
     stats: BatchStats
+    caches: Optional[List[List[Any]]]
 
+def _left_pad_prompts(prompts: Any, max_length: Optional[int] = ...) -> mx.array: ...
+def _right_pad_prompts(prompts: Any, max_length: Optional[int] = ...) -> mx.array: ...
+def _make_cache(
+    model: Any, left_padding: Any, max_kv_size: Optional[int]
+) -> List[Any]: ...
+def _merge_caches(caches: Any) -> List[Any]: ...
 @dataclass
 class Batch:
     uids: List[int]
@@ -261,39 +270,71 @@ class Batch:
     max_tokens: List[int]
     num_tokens: List[int]
     cache: List[Any]
-    def __len__(self):  # -> int:
-        ...
-    def filter(self, keep_idx: List[int]):  # -> None:
-        ...
-    def extend(self, other):  # -> None:
-        ...
+    samplers: List[Any]
+    logits_processors: List[Any]
+    tokens: List[mx.array]
+    def __len__(self) -> int: ...
+    def filter(self, keep_idx: List[int]) -> None: ...
+    def extend(self, other: "Batch") -> None: ...
+    def extract_cache(self, idx: int) -> List[Any]: ...
 
 class BatchGenerator:
+    model: Any
+    max_kv_size: Optional[int]
+    prefill_step_size: int
+    unprocessed_prompts: List[Any]
+    active_batch: Optional[Batch]
+    prompt_progress_callback: Callable[[List[Tuple[int, int, int]]], None]
+    _stats: BatchStats
+
     @dataclass
     class Response:
         uid: int
         token: int
         logprobs: mx.array
         finish_reason: Optional[str]
+        prompt_cache: Any
 
     def __init__(
         self,
-        model,
+        model: Any,
         max_tokens: int = ...,
-        stop_tokens: Optional[set] = ...,
+        stop_tokens: Optional[set[int]] = ...,
         sampler: Optional[Callable[[mx.array], mx.array]] = ...,
+        logits_processors: Optional[
+            List[Callable[[mx.array, mx.array], mx.array]]
+        ] = ...,
         completion_batch_size: int = ...,
         prefill_batch_size: int = ...,
         prefill_step_size: int = ...,
+        prompt_progress_callback: Optional[
+            Callable[[List[Tuple[int, int, int]]], None]
+        ] = ...,
+        max_kv_size: Optional[int] = ...,
     ) -> None: ...
+    def close(self) -> None: ...
     def insert(
-        self, prompts, max_tokens: Union[List[int], int, None] = ...
-    ):  # -> list[Any]:
-        ...
-    def stats(self):  # -> BatchStats:
-        ...
-    def next(self):  # -> list[Any]:
-        ...
+        self,
+        prompts: Any,
+        max_tokens: Union[List[int], int, None] = ...,
+        caches: Any = ...,
+        samplers: Optional[List[Any]] = ...,
+        logits_processors: Optional[List[Any]] = ...,
+    ) -> List[int]: ...
+    def remove(
+        self, uids: List[int], return_prompt_caches: bool = ...
+    ) -> Optional[dict[int, List[Any]]]: ...
+    def stats(self) -> BatchStats: ...
+    def next(self) -> List[Response]: ...
+    def _process_prompts(self, prompts: List[Any]) -> Batch: ...
+    def _step(
+        self,
+        input_tokens: mx.array,
+        prompt_cache: List[Any],
+        samplers: Optional[List[Any]],
+        logits_processors: Optional[List[Any]],
+        tokens: List[mx.array],
+    ) -> Tuple[mx.array, List[mx.array]]: ...
 
 def batch_generate(
     model,

.mlx_typings/mlx_lm/models/cache.pyi

@@ -16,7 +16,7 @@ class Cache(Protocol):
         self, keys: mx.array, values: mx.array
     ) -> tuple[mx.array, mx.array]: ...
     @property
-    def state(self) -> tuple[mx.array, mx.array]: ...
+    def state(self) -> tuple[mx.array | None, mx.array | None]: ...
     @state.setter
     def state(self, v) -> None: ...
 
@@ -92,13 +92,14 @@ class _BaseCache(Cache):
     values: mx.array
     offset: int
     @property
-    def state(self) -> tuple[mx.array, mx.array]: ...
+    def state(self) -> tuple[mx.array | None, mx.array | None]: ...
     @state.setter
     def state(self, v) -> None: ...
     @property
     def meta_state(self) -> Literal[""]: ...
     @meta_state.setter
     def meta_state(self, v) -> None: ...
+    def trim(self, n: int) -> int: ...
     def is_trimmable(self) -> Literal[False]: ...
     @classmethod
     def from_state(cls, state, meta_state) -> Self: ...
@@ -114,15 +115,13 @@ class ConcatenateKVCache(_BaseCache):
     def update_and_fetch(self, keys, values):  # -> tuple[Any | array, Any | array]:
         ...
     @property
-    def state(self):  # -> tuple[Any | array | None, Any | array | None]:
-        ...
+    def state(self) -> tuple[mx.array | None, mx.array | None]: ...
     @state.setter
     def state(self, v):  # -> None:
         ...
     def is_trimmable(self):  # -> Literal[True]:
         ...
-    def trim(self, n):  # -> int:
-        ...
+    def trim(self, n: int) -> int: ...
     def make_mask(self, *args, **kwargs):  # -> array | Literal['causal'] | None:
         ...
 
@@ -132,10 +131,7 @@ class QuantizedKVCache(_BaseCache):
     def update_and_fetch(self, keys, values):  # -> Any:
         ...
     @property
-    def state(
-        self,
-    ):  # -> tuple[Any | tuple[array, array, array] | None, Any | tuple[array, array, array] | None] | Any:
-        ...
+    def state(self) -> tuple[mx.array | None, mx.array | None]: ...
     @state.setter
     def state(self, v):  # -> None:
         ...
@@ -147,8 +143,7 @@ class QuantizedKVCache(_BaseCache):
         ...
     def is_trimmable(self):  # -> Literal[True]:
         ...
-    def trim(self, n):  # -> int:
-        ...
+    def trim(self, n: int) -> int: ...
     def make_mask(self, *args, **kwargs):  # -> array | Literal['causal'] | None:
         ...
 
@@ -160,22 +155,30 @@ class KVCache(_BaseCache):
     @property
     def state(
         self,
-    ) -> tuple[array, array]: ...
+    ) -> tuple[mx.array | None, mx.array | None]: ...
     @state.setter
     def state(self, v) -> None: ...
     def is_trimmable(self):  # -> Literal[True]:
         ...
-    def trim(self, n):  # -> int:
-        ...
+    def trim(self, n: int) -> int: ...
     def to_quantized(
         self, group_size: int = ..., bits: int = ...
     ) -> QuantizedKVCache: ...
-    def make_mask(self, *args, **kwargs):  # -> array | Literal['causal'] | None:
-        ...
+    def make_mask(
+        self, *args: Any, **kwargs: Any
+    ) -> mx.array | Literal["causal"] | None: ...
 
 class RotatingKVCache(_BaseCache):
     step = ...
+    keys: mx.array | None
+    values: mx.array | None
+    keep: int
+    max_size: int
+    _idx: int
     def __init__(self, max_size, keep=...) -> None: ...
+    def _trim(
+        self, trim_size: int, v: mx.array, append: mx.array | None = ...
+    ) -> mx.array: ...
     def update_and_fetch(
         self, keys, values
     ):  # -> tuple[array | Any, array | Any] | tuple[array | Any, array | Any | None]:
@@ -183,8 +186,7 @@ class RotatingKVCache(_BaseCache):
     @property
     def state(
         self,
-    ):  # -> tuple[Any | array, Any | array] | tuple[Any | array | None, Any | array | None]:
-        ...
+    ) -> tuple[mx.array | None, mx.array | None]: ...
     @state.setter
     def state(self, v):  # -> None:
         ...
@@ -196,8 +198,7 @@ class RotatingKVCache(_BaseCache):
         ...
     def is_trimmable(self):  # -> bool:
         ...
-    def trim(self, n):  # -> int:
-        ...
+    def trim(self, n: int) -> int: ...
     def to_quantized(
         self, group_size: int = ..., bits: int = ...
     ) -> QuantizedKVCache: ...
@@ -212,8 +213,7 @@ class ArraysCache(_BaseCache):
         ...
     def __getitem__(self, idx): ...
     @property
-    def state(self):  # -> list[Any | array] | list[array]:
-        ...
+    def state(self) -> tuple[mx.array | None, mx.array | None]: ...
     @state.setter
     def state(self, v):  # -> None:
         ...
@@ -227,8 +227,7 @@ class ArraysCache(_BaseCache):
         In-place extend this cache with the other cache.
         """
 
-    def make_mask(self, N: int):  # -> array | None:
-        ...
+    def make_mask(self, N: int) -> mx.array | None: ...
 
 class MambaCache(ArraysCache):
     def __init__(self, left_padding: Optional[List[int]] = ...) -> None: ...
@@ -239,8 +238,7 @@ class ChunkedKVCache(KVCache):
         ...
     def update_and_fetch(self, keys, values):  # -> tuple[array, array]:
         ...
-    def trim(self, n):  # -> int:
-        ...
+    def trim(self, n: int) -> int: ...
     @property
     def meta_state(self):  # -> tuple[str, ...]:
         ...
@@ -253,10 +251,9 @@ class CacheList(_BaseCache):
     def __getitem__(self, idx): ...
     def is_trimmable(self):  # -> bool:
         ...
-    def trim(self, n): ...
+    def trim(self, n: int) -> int: ...
     @property
-    def state(self):  # -> list[Any]:
-        ...
+    def state(self) -> list[tuple[mx.array | None, mx.array | None]]: ...
     @state.setter
     def state(self, v):  # -> None:
         ...

.mlx_typings/mlx_lm/models/nemotron_h.pyi

@@ -0,0 +1,154 @@
+from dataclasses import dataclass
+from typing import Any, List, Optional, Tuple
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .cache import ArraysCache, KVCache
+from .switch_layers import SwitchMLP
+
+@dataclass
+class ModelArgs:
+    model_type: str
+    vocab_size: int
+    hidden_size: int
+    intermediate_size: int
+    num_hidden_layers: int
+    max_position_embeddings: int
+    num_attention_heads: int
+    num_key_value_heads: int
+    attention_bias: bool
+    mamba_num_heads: int
+    mamba_head_dim: int
+    mamba_proj_bias: bool
+    ssm_state_size: int
+    conv_kernel: int
+    n_groups: int
+    mlp_bias: bool
+    layer_norm_epsilon: float
+    use_bias: bool
+    use_conv_bias: bool
+    hybrid_override_pattern: List[str]
+    head_dim: Optional[int]
+    moe_intermediate_size: Optional[int]
+    moe_shared_expert_intermediate_size: Optional[int]
+    n_group: Optional[int]
+    n_routed_experts: Optional[int]
+    n_shared_experts: Optional[int]
+    topk_group: Optional[int]
+    num_experts_per_tok: Optional[int]
+    norm_topk_prob: Optional[bool]
+    routed_scaling_factor: Optional[float]
+    time_step_limit: Optional[Tuple[float, float]]
+    time_step_min: Optional[float]
+    time_step_max: Optional[float]
+
+    @classmethod
+    def from_dict(cls, params: dict[str, Any]) -> ModelArgs: ...
+    def __post_init__(self) -> None: ...
+
+class NemotronHMamba2Mixer(nn.Module):
+    num_heads: int
+    hidden_size: int
+    ssm_state_size: int
+    conv_kernel_size: int
+    intermediate_size: int
+    n_groups: int
+    head_dim: int
+    conv_dim: int
+    conv1d: nn.Conv1d
+    in_proj: nn.Linear
+    dt_bias: mx.array
+    A_log: mx.array
+    D: mx.array
+    norm: nn.RMSNorm
+    heads_per_group: int
+    out_proj: nn.Linear
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(
+        self,
+        hidden_states: mx.array,
+        mask: Optional[mx.array],
+        cache: Optional[ArraysCache] = None,
+    ) -> mx.array: ...
+
+class NemotronHAttention(nn.Module):
+    hidden_size: int
+    num_heads: int
+    head_dim: int
+    num_key_value_heads: int
+    scale: float
+    q_proj: nn.Linear
+    k_proj: nn.Linear
+    v_proj: nn.Linear
+    o_proj: nn.Linear
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[KVCache] = None,
+    ) -> mx.array: ...
+
+class NemotronHMLP(nn.Module):
+    up_proj: nn.Linear
+    down_proj: nn.Linear
+
+    def __init__(
+        self, args: ModelArgs, intermediate_size: Optional[int] = None
+    ) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class NemotronHMoE(nn.Module):
+    num_experts_per_tok: int
+    switch_mlp: SwitchMLP
+    shared_experts: NemotronHMLP
+
+    def __init__(self, config: ModelArgs) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class NemotronHBlock(nn.Module):
+    block_type: str
+    norm: nn.RMSNorm
+    mixer: NemotronHMamba2Mixer | NemotronHAttention | NemotronHMLP | NemotronHMoE
+
+    def __init__(self, args: ModelArgs, block_type: str) -> None: ...
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+
+class NemotronHModel(nn.Module):
+    embeddings: nn.Embedding
+    layers: list[NemotronHBlock]
+    norm_f: nn.RMSNorm
+    fa_idx: int
+    ssm_idx: int
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+
+class Model(nn.Module):
+    args: ModelArgs
+    backbone: NemotronHModel
+    lm_head: nn.Linear
+    model_type: str
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+    @property
+    def layers(self) -> list[NemotronHBlock]: ...
+    def make_cache(self) -> list[ArraysCache | KVCache]: ...
+    def sanitize(self, weights: dict[str, Any]) -> dict[str, Any]: ...

.mlx_typings/mlx_lm/models/qwen3_5.pyi

@@ -0,0 +1,153 @@
+from dataclasses import dataclass
+from typing import Any, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .cache import ArraysCache, KVCache
+from .qwen3_next import (
+    Qwen3NextAttention as Attention,
+    Qwen3NextMLP as MLP,
+    Qwen3NextRMSNormGated as RMSNormGated,
+    Qwen3NextSparseMoeBlock,
+)
+
+SparseMoeBlock = Qwen3NextSparseMoeBlock
+from .switch_layers import SwitchGLU
+
+@dataclass
+class TextModelArgs:
+    model_type: str
+    hidden_size: int
+    intermediate_size: int
+    num_hidden_layers: int
+    num_attention_heads: int
+    rms_norm_eps: float
+    vocab_size: int
+    num_key_value_heads: int
+    max_position_embeddings: int
+    linear_num_value_heads: int
+    linear_num_key_heads: int
+    linear_key_head_dim: int
+    linear_value_head_dim: int
+    linear_conv_kernel_dim: int
+    tie_word_embeddings: bool
+    attention_bias: bool
+    head_dim: Optional[int]
+    full_attention_interval: int
+    num_experts: int
+    num_experts_per_tok: int
+    decoder_sparse_step: int
+    shared_expert_intermediate_size: int
+    moe_intermediate_size: int
+    norm_topk_prob: bool
+    rope_parameters: Optional[dict[str, Any]]
+    partial_rotary_factor: float
+    rope_theta: float
+    rope_scaling: Optional[dict[str, Any]]
+
+    @classmethod
+    def from_dict(cls, params: dict[str, Any]) -> TextModelArgs: ...
+    def __post_init__(self) -> None: ...
+
+class GatedDeltaNet(nn.Module):
+    hidden_size: int
+    num_v_heads: int
+    num_k_heads: int
+    head_k_dim: int
+    head_v_dim: int
+    key_dim: int
+    value_dim: int
+    conv_kernel_size: int
+    conv_dim: int
+    conv1d: nn.Conv1d
+    in_proj_qkv: nn.Linear
+    in_proj_z: nn.Linear
+    in_proj_b: nn.Linear
+    in_proj_a: nn.Linear
+    dt_bias: mx.array
+    A_log: mx.array
+    norm: RMSNormGated
+    out_proj: nn.Linear
+
+    def __init__(self, config: TextModelArgs) -> None: ...
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+
+class DecoderLayer(nn.Module):
+    is_linear: bool
+    linear_attn: GatedDeltaNet
+    self_attn: Attention
+    input_layernorm: nn.RMSNorm
+    post_attention_layernorm: nn.RMSNorm
+    mlp: MLP | SparseMoeBlock
+
+    def __init__(self, args: TextModelArgs, layer_idx: int) -> None: ...
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+
+class Qwen3_5TextModel(nn.Module):
+    embed_tokens: nn.Embedding
+    layers: list[DecoderLayer]
+    norm: nn.RMSNorm
+    ssm_idx: int
+    fa_idx: int
+
+    def __init__(self, args: TextModelArgs) -> None: ...
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+        input_embeddings: Optional[mx.array] = None,
+    ) -> mx.array: ...
+
+class TextModel(nn.Module):
+    args: TextModelArgs
+    model_type: str
+    model: Qwen3_5TextModel
+    lm_head: nn.Linear
+
+    def __init__(self, args: TextModelArgs) -> None: ...
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+        input_embeddings: Optional[mx.array] = None,
+    ) -> mx.array: ...
+    @property
+    def layers(self) -> list[DecoderLayer]: ...
+    def make_cache(self) -> list[ArraysCache | KVCache]: ...
+    def sanitize(self, weights: dict[str, Any]) -> dict[str, Any]: ...
+
+@dataclass
+class ModelArgs:
+    model_type: str
+    text_config: dict[str, Any]
+
+    @classmethod
+    def from_dict(cls, params: dict[str, Any]) -> ModelArgs: ...
+
+class Model(nn.Module):
+    args: ModelArgs
+    model_type: str
+    language_model: TextModel
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+        input_embeddings: Optional[mx.array] = None,
+    ) -> mx.array: ...
+    def sanitize(self, weights: dict[str, Any]) -> dict[str, Any]: ...
+    @property
+    def layers(self) -> list[DecoderLayer]: ...
+    def make_cache(self) -> list[ArraysCache | KVCache]: ...

.mlx_typings/mlx_lm/models/qwen3_5_moe.pyi

@@ -0,0 +1,19 @@
+from dataclasses import dataclass
+from typing import Any, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .cache import ArraysCache, KVCache
+from .qwen3_5 import DecoderLayer, Model as Qwen3_5Model, TextModel
+
+@dataclass
+class ModelArgs:
+    model_type: str
+    text_config: dict[str, Any]
+
+    @classmethod
+    def from_dict(cls, params: dict[str, Any]) -> ModelArgs: ...
+
+class Model(Qwen3_5Model):
+    def sanitize(self, weights: dict[str, Any]) -> dict[str, Any]: ...

.mlx_typings/mlx_lm/models/qwen3_next.pyi

@@ -5,8 +5,18 @@ from typing import Any, Optional
 import mlx.core as mx
 import mlx.nn as nn
 
+from .cache import ArraysCache, KVCache
 from .switch_layers import SwitchGLU
 
+class Qwen3NextRMSNormGated(nn.Module):
+    eps: float
+    weight: mx.array
+
+    def __init__(self, hidden_size: int, eps: float = ...) -> None: ...
+    def __call__(
+        self, hidden_states: mx.array, gate: mx.array | None = None
+    ) -> mx.array: ...
+
 class Qwen3NextMLP(nn.Module):
     gate_proj: nn.Linear
     down_proj: nn.Linear
@@ -90,6 +100,8 @@ class Qwen3NextModel(nn.Module):
     embed_tokens: nn.Embedding
     layers: list[Qwen3NextDecoderLayer]
     norm: nn.RMSNorm
+    ssm_idx: int
+    fa_idx: int
 
     def __init__(self, args: Any) -> None: ...
     def __call__(
@@ -112,3 +124,4 @@ class Model(nn.Module):
     def sanitize(self, weights: dict[str, Any]) -> dict[str, Any]: ...
     @property
     def layers(self) -> list[Qwen3NextDecoderLayer]: ...
+    def make_cache(self) -> list[ArraysCache | KVCache]: ...

.mlx_typings/mlx_lm/models/switch_layers.pyi

@@ -73,6 +73,9 @@ class SwitchGLU(nn.Module):
     def __call__(self, x, indices) -> mx.array: ...
 
 class SwitchMLP(nn.Module):
+    fc1: SwitchLinear
+    fc2: SwitchLinear
+
     def __init__(
         self,
         input_dims: int,

.mlx_typings/mlx_lm/sample_utils.pyi

@@ -48,7 +48,7 @@ def make_logits_processors(
     logit_bias: Optional[Dict[int, float]] = ...,
     repetition_penalty: Optional[float] = ...,
     repetition_context_size: Optional[int] = ...,
-):  # -> list[Any]:
+) -> list[Callable[[mx.array, mx.array], mx.array]]:
     """
     Make logits processors for use with ``generate_step``.
 

.mlx_typings/mlx_lm/tokenizer_utils.pyi

@@ -39,11 +39,11 @@ class StreamingDetokenizer:
     """
 
     __slots__ = ...
-    def reset(self): ...
-    def add_token(self, token): ...
-    def finalize(self): ...
+    def reset(self) -> None: ...
+    def add_token(self, token: int) -> None: ...
+    def finalize(self) -> None: ...
     @property
-    def last_segment(self):
+    def last_segment(self) -> str:
         """Return the last segment of readable text since last time this property was accessed."""
 
 class NaiveStreamingDetokenizer(StreamingDetokenizer):

CONTRIBUTING.md

@@ -39,6 +39,119 @@ Write pure functions where possible. When adding new code, prefer Rust unless th
 
 Run `nix fmt` to auto-format your code before submitting.
 
+## Model Cards
+
+EXO uses TOML-based model cards to define model metadata and capabilities. Model cards are stored in:
+- `resources/inference_model_cards/` for text generation models
+- `resources/image_model_cards/` for image generation models
+- `~/.exo/custom_model_cards/` for user-added custom models
+
+### Adding a Model Card
+
+To add a new model, create a TOML file with the following structure:
+
+```toml
+model_id = "mlx-community/Llama-3.2-1B-Instruct-4bit"
+n_layers = 16
+hidden_size = 2048
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "llama"
+quantization = "4bit"
+base_model = "Llama 3.2 1B"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 729808896
+```
+
+### Required Fields
+
+- `model_id`: Hugging Face model identifier
+- `n_layers`: Number of transformer layers
+- `hidden_size`: Hidden dimension size
+- `supports_tensor`: Whether the model supports tensor parallelism
+- `tasks`: List of supported tasks (`TextGeneration`, `TextToImage`, `ImageToImage`)
+- `family`: Model family (e.g., "llama", "deepseek", "qwen")
+- `quantization`: Quantization level (e.g., "4bit", "8bit", "bf16")
+- `base_model`: Human-readable base model name
+- `capabilities`: List of capabilities (e.g., `["text"]`, `["text", "thinking"]`)
+
+### Optional Fields
+
+- `components`: For multi-component models (like image models with separate text encoders and transformers)
+- `uses_cfg`: Whether the model uses classifier-free guidance (for image models)
+- `trust_remote_code`: Whether to allow remote code execution (defaults to `false` for security)
+
+### Capabilities
+
+The `capabilities` field defines what the model can do:
+- `text`: Standard text generation
+- `thinking`: Model supports chain-of-thought reasoning
+- `thinking_toggle`: Thinking can be enabled/disabled via `enable_thinking` parameter
+- `image_edit`: Model supports image-to-image editing (FLUX.1-Kontext)
+
+### Security Note
+
+By default, `trust_remote_code` is set to `false` for security. Only enable it if the model explicitly requires remote code execution from the Hugging Face hub.
+
+## API Adapters
+
+EXO supports multiple API formats through an adapter pattern. Adapters convert API-specific request formats to the internal `TextGenerationTaskParams` format and convert internal token chunks back to API-specific responses.
+
+### Adapter Architecture
+
+All adapters live in `src/exo/master/adapters/` and follow the same pattern:
+
+1. Convert API-specific requests to `TextGenerationTaskParams`
+2. Handle both streaming and non-streaming response generation
+3. Convert internal `TokenChunk` objects to API-specific formats
+4. Manage error handling and edge cases
+
+### Existing Adapters
+
+- `chat_completions.py`: OpenAI Chat Completions API
+- `claude.py`: Anthropic Claude Messages API
+- `responses.py`: OpenAI Responses API
+- `ollama.py`: Ollama API (for OpenWebUI compatibility)
+
+### Adding a New API Adapter
+
+To add support for a new API format:
+
+1. Create a new adapter file in `src/exo/master/adapters/`
+2. Implement a request conversion function:
+   ```python
+   def your_api_request_to_text_generation(
+       request: YourAPIRequest,
+   ) -> TextGenerationTaskParams:
+       # Convert API request to internal format
+       pass
+   ```
+3. Implement streaming response generation:
+   ```python
+   async def generate_your_api_stream(
+       command_id: CommandId,
+       chunk_stream: AsyncGenerator[TokenChunk | ErrorChunk | ToolCallChunk, None],
+   ) -> AsyncGenerator[str, None]:
+       # Convert internal chunks to API-specific streaming format
+       pass
+   ```
+4. Implement non-streaming response collection:
+   ```python
+   async def collect_your_api_response(
+       command_id: CommandId,
+       chunk_stream: AsyncGenerator[TokenChunk | ErrorChunk | ToolCallChunk, None],
+   ) -> AsyncGenerator[str]:
+       # Collect all chunks and return single response
+       pass
+   ```
+5. Register the adapter endpoints in `src/exo/master/api.py`
+
+The adapter pattern keeps API-specific logic isolated from core inference systems. Internal systems (worker, runner, event sourcing) only see `TextGenerationTaskParams` and `TokenChunk` objects - no API-specific types cross the adapter boundary.
+
+For detailed API documentation, see [docs/api.md](docs/api.md).
+
 ## Testing
 
 EXO relies heavily on manual testing at this point in the project, but this is evolving. Before submitting a change, test both before and after to demonstrate how your change improves behavior. Do the best you can with the hardware you have available - if you need help testing, ask and we'll do our best to assist. Add automated tests where possible - we're actively working to substantially improve our automated testing story.

Cargo.lock

@@ -216,6 +216,28 @@ dependencies = [
  "windows-sys 0.61.2",
 ]
 
+[[package]]
+name = "async-stream"
+version = "0.3.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0b5a71a6f37880a80d1d7f19efd781e4b5de42c88f0722cc13bcb6cc2cfe8476"
+dependencies = [
+ "async-stream-impl",
+ "futures-core",
+ "pin-project-lite",
+]
+
+[[package]]
+name = "async-stream-impl"
+version = "0.3.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c7c24de15d275a1ecfd47a380fb4d5ec9bfe0933f309ed5e705b775596a3574d"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.111",
+]
+
 [[package]]
 name = "async-trait"
 version = "0.1.89"
@@ -2759,6 +2781,7 @@ dependencies = [
 name = "networking"
 version = "0.0.1"
 dependencies = [
+ "async-stream",
  "delegate",
  "either",
  "extend",
@@ -2767,6 +2790,7 @@ dependencies = [
  "keccak-const",
  "libp2p",
  "log",
+ "pin-project",
  "tokio",
  "tracing-subscriber",
  "util",

Cargo.toml

@@ -1,10 +1,6 @@
 [workspace]
 resolver = "3"
-members = [
-    "rust/networking",
-    "rust/exo_pyo3_bindings",
-    "rust/util",
-]
+members = ["rust/networking", "rust/exo_pyo3_bindings", "rust/util"]
 
 [workspace.package]
 version = "0.0.1"
@@ -34,6 +30,7 @@ delegate = "0.13"
 keccak-const = "0.2"
 
 # Async dependencies
+async-stream = "0.3"
 tokio = "1.46"
 futures-lite = "2.6.1"
 futures-timer = "3.0"

README.md

@@ -26,6 +26,8 @@ exo connects all your devices into an AI cluster. Not only does exo enable runni
 - **Topology-Aware Auto Parallel**: exo figures out the best way to split your model across all available devices based on a realtime view of your device topology. It takes into account device resources and network latency/bandwidth between each link.
 - **Tensor Parallelism**: exo supports sharding models, for up to 1.8x speedup on 2 devices and 3.2x speedup on 4 devices.
 - **MLX Support**: exo uses [MLX](https://github.com/ml-explore/mlx) as an inference backend and [MLX distributed](https://ml-explore.github.io/mlx/build/html/usage/distributed.html) for distributed communication.
+- **Multiple API Compatibility**: Compatible with OpenAI Chat Completions API, Claude Messages API, OpenAI Responses API, and Ollama API - use your existing tools and clients.
+- **Custom Model Support**: Load custom models from HuggingFace hub to expand the range of available models.
 
 ## Dashboard
 
@@ -196,6 +198,8 @@ exo follows the [XDG Base Directory Specification](https://specifications.freede
 - **Configuration files**: `~/.config/exo/` (or `$XDG_CONFIG_HOME/exo/`)
 - **Data files**: `~/.local/share/exo/` (or `$XDG_DATA_HOME/exo/`)
 - **Cache files**: `~/.cache/exo/` (or `$XDG_CACHE_HOME/exo/`)
+- **Log files**: `~/.cache/exo/exo_log/` (with automatic log rotation)
+- **Custom model cards**: `~/.local/share/exo/custom_model_cards/`
 
 You can override these locations by setting the corresponding XDG environment variables.
 
@@ -275,8 +279,47 @@ After that, RDMA will be enabled in macOS and exo will take care of the rest.
 
 ---
 
+## Environment Variables
+
+exo supports several environment variables for configuration:
+
+| Variable | Description | Default |
+|----------|-------------|---------|
+| `EXO_MODELS_PATH` | Colon-separated paths to search for pre-downloaded models (e.g., on NFS mounts or shared storage) | None |
+| `EXO_MODELS_DIR` | Directory where exo downloads and stores models | `~/.local/share/exo/models` (Linux) or `~/.exo/models` (macOS) |
+| `EXO_OFFLINE` | Run without internet connection (uses only local models) | `false` |
+| `EXO_ENABLE_IMAGE_MODELS` | Enable image model support | `false` |
+| `EXO_LIBP2P_NAMESPACE` | Custom namespace for cluster isolation | None |
+| `EXO_FAST_SYNCH` | Control MLX_METAL_FAST_SYNCH behavior (for JACCL backend) | Auto |
+| `EXO_TRACING_ENABLED` | Enable distributed tracing for performance analysis | `false` |
+
+**Example usage:**
+
+```bash
+# Use pre-downloaded models from NFS mount
+EXO_MODELS_PATH=/mnt/nfs/models:/opt/ai-models uv run exo
+
+# Run in offline mode
+EXO_OFFLINE=true uv run exo
+
+# Enable image models
+EXO_ENABLE_IMAGE_MODELS=true uv run exo
+
+# Use custom namespace for cluster isolation
+EXO_LIBP2P_NAMESPACE=my-dev-cluster uv run exo
+```
+
+---
+
 ### Using the API
 
+exo provides multiple API-compatible interfaces for maximum compatibility with existing tools:
+
+- **OpenAI Chat Completions API** - Compatible with OpenAI clients
+- **Claude Messages API** - Compatible with Anthropic's Claude format
+- **OpenAI Responses API** - Compatible with OpenAI's Responses format
+- **Ollama API** - Compatible with Ollama and tools like OpenWebUI
+
 If you prefer to interact with exo via the API, here is an example creating an instance of a small model (`mlx-community/Llama-3.2-1B-Instruct-4bit`), sending a chat completions request and deleting the instance.
 
 ---
@@ -366,14 +409,85 @@ When you're done, delete the instance by its ID (find it via `/state` or `/insta
 curl -X DELETE http://localhost:52415/instance/YOUR_INSTANCE_ID
 ```
 
+### Claude Messages API Compatibility
+
+Use the Claude Messages API format with the `/v1/messages` endpoint:
+
+```bash
+curl -N -X POST http://localhost:52415/v1/messages \
+  -H 'Content-Type: application/json' \
+  -d '{
+    "model": "mlx-community/Llama-3.2-1B-Instruct-4bit",
+    "messages": [
+      {"role": "user", "content": "Hello"}
+    ],
+    "max_tokens": 1024,
+    "stream": true
+  }'
+```
+
+### OpenAI Responses API Compatibility
+
+Use the OpenAI Responses API format with the `/v1/responses` endpoint:
+
+```bash
+curl -N -X POST http://localhost:52415/v1/responses \
+  -H 'Content-Type: application/json' \
+  -d '{
+    "model": "mlx-community/Llama-3.2-1B-Instruct-4bit",
+    "messages": [
+      {"role": "user", "content": "Hello"}
+    ],
+    "stream": true
+  }'
+```
+
+### Ollama API Compatibility
+
+exo supports Ollama API endpoints for compatibility with tools like OpenWebUI:
+
+```bash
+# Ollama chat
+curl -X POST http://localhost:52415/ollama/api/chat \
+  -H 'Content-Type: application/json' \
+  -d '{
+    "model": "mlx-community/Llama-3.2-1B-Instruct-4bit",
+    "messages": [
+      {"role": "user", "content": "Hello"}
+    ],
+    "stream": false
+  }'
+
+# List models (Ollama format)
+curl http://localhost:52415/ollama/api/tags
+```
+
+### Custom Model Loading from HuggingFace
+
+You can add custom models from the HuggingFace hub:
+
+```bash
+curl -X POST http://localhost:52415/models/add \
+  -H 'Content-Type: application/json' \
+  -d '{
+    "model_id": "mlx-community/my-custom-model"
+  }'
+```
+
+**Security Note:**
+
+Custom models requiring `trust_remote_code` in their configuration must be explicitly enabled (default is false) for security. Only enable this if you trust the model's remote code execution. Models are fetched from HuggingFace and stored locally as custom model cards.
+
 **Other useful API endpoints*:**
 
 - List all models: `curl http://localhost:52415/models`
+- List downloaded models only: `curl http://localhost:52415/models?status=downloaded`
+- Search HuggingFace: `curl "http://localhost:52415/models/search?query=llama&limit=10"`
 - Inspect instance IDs and deployment state: `curl http://localhost:52415/state`
 
 For further details, see:
 
-- API basic documentation in [docs/api.md](docs/api.md).
+- API documentation in [docs/api.md](docs/api.md).
 - API types and endpoints in [src/exo/master/api.py](src/exo/master/api.py).
 
 ---
@@ -432,4 +546,4 @@ On macOS, exo uses the GPU. On Linux, exo currently runs on CPU. We are working
 
 ## Contributing
 
-See [CONTRIBUTING.md](CONTRIBUTING.md) for guidelines on how to contribute to exo.
+See [CONTRIBUTING.md](CONTRIBUTING.md) for guidelines on how to contribute to exo.

bench/bench.toml

@@ -2,6 +2,4 @@
 #
 # Lists the suite files to include. Each file defines benchmarks
 # with shared constraints, topology, and default args.
-include = [
-    "single-m3-ultra.toml",
-]
+include = ["single-m3-ultra.toml"]

bench/eval_configs/models.toml

@@ -0,0 +1,292 @@
+# Model evaluation configurations for exo_eval.
+#
+# Each [[model]] entry uses `patterns` — a list of substrings matched
+# against the model_id. First matching entry wins.
+#
+# Required fields:
+#   name, patterns, reasoning
+#
+# Optional per-model overrides (CLI flags take priority over these):
+#   temperature, top_p, max_tokens, reasoning_effort
+#
+# Fallback defaults (when no per-model config):
+#   reasoning:     temperature=1.0, max_tokens=131072, reasoning_effort="high"
+#   non-reasoning: temperature=0.0, max_tokens=16384
+#
+# All per-model values below are sourced from official model cards,
+# generation_config.json files, and vendor documentation.
+
+# ─── Qwen3.5 (Feb 2026) ─────────────────────────────────────────────
+# Source: HuggingFace model cards (Qwen/Qwen3.5-*)
+# 35B-A3B thinking general: temp=1.0, top_p=0.95, top_k=20
+# 397B thinking: temp=0.6, top_p=0.95, top_k=20
+# Non-thinking: temp=0.7, top_p=0.8, top_k=20
+# max_tokens: 32768 general, 81920 for complex math/code
+
+[[model]]
+name = "Qwen3.5 2B"
+patterns = ["Qwen3.5-2B"]
+reasoning = true
+temperature = 0.6
+top_p = 0.95
+max_tokens = 81920
+
+[[model]]
+name = "Qwen3.5 9B"
+patterns = ["Qwen3.5-9B"]
+reasoning = true
+temperature = 0.6
+top_p = 0.95
+max_tokens = 81920
+
+[[model]]
+name = "Qwen3.5 27B"
+patterns = ["Qwen3.5-27B"]
+reasoning = true
+temperature = 0.6
+top_p = 0.95
+max_tokens = 81920
+
+[[model]]
+name = "Qwen3.5 35B A3B"
+patterns = ["Qwen3.5-35B-A3B"]
+reasoning = true
+temperature = 1.0
+top_p = 0.95
+max_tokens = 81920
+
+[[model]]
+name = "Qwen3.5 122B A10B"
+patterns = ["Qwen3.5-122B-A10B"]
+reasoning = true
+temperature = 0.6
+top_p = 0.95
+max_tokens = 81920
+
+[[model]]
+name = "Qwen3.5 397B A17B"
+patterns = ["Qwen3.5-397B-A17B"]
+reasoning = true
+temperature = 0.6
+top_p = 0.95
+max_tokens = 81920
+
+# ─── Qwen3 (Apr 2025) ───────────────────────────────────────────────
+# Source: HuggingFace model cards (Qwen/Qwen3-*)
+# Thinking: temp=0.6, top_p=0.95, top_k=20
+# Non-thinking: temp=0.7, top_p=0.8, top_k=20
+# max_tokens: 32768 general, 38912 for complex math/code
+
+[[model]]
+name = "Qwen3 0.6B"
+patterns = ["Qwen3-0.6B"]
+reasoning = true
+temperature = 0.6
+top_p = 0.95
+max_tokens = 38912
+
+[[model]]
+name = "Qwen3 30B A3B"
+patterns = ["Qwen3-30B-A3B"]
+reasoning = true
+temperature = 0.6
+top_p = 0.95
+max_tokens = 38912
+
+[[model]]
+name = "Qwen3 235B A22B"
+patterns = ["Qwen3-235B-A22B"]
+reasoning = true
+temperature = 0.6
+top_p = 0.95
+max_tokens = 38912
+
+[[model]]
+name = "Qwen3 Next 80B Thinking"
+patterns = ["Qwen3-Next-80B-A3B-Thinking"]
+reasoning = true
+temperature = 0.6
+top_p = 0.95
+max_tokens = 38912
+
+[[model]]
+name = "Qwen3 Next 80B Instruct"
+patterns = ["Qwen3-Next-80B-A3B-Instruct"]
+reasoning = false
+temperature = 0.7
+top_p = 0.8
+max_tokens = 16384
+
+[[model]]
+name = "Qwen3 Coder 480B"
+patterns = ["Qwen3-Coder-480B"]
+reasoning = false
+temperature = 0.7
+top_p = 0.8
+max_tokens = 16384
+
+[[model]]
+name = "Qwen3 Coder Next"
+patterns = ["Qwen3-Coder-Next"]
+reasoning = false
+temperature = 0.7
+top_p = 0.8
+max_tokens = 16384
+
+# ─── GPT-OSS (OpenAI) ───────────────────────────────────────────────
+# Source: OpenAI GitHub README + HuggingFace discussion #21
+# temp=1.0, top_p=1.0, NO top_k, NO repetition_penalty
+# reasoning_effort supported: low/medium/high
+
+[[model]]
+name = "GPT-OSS 20B"
+patterns = ["gpt-oss-20b"]
+reasoning = true
+temperature = 1.0
+top_p = 1.0
+
+[[model]]
+name = "GPT-OSS 120B"
+patterns = ["gpt-oss-120b"]
+reasoning = true
+temperature = 1.0
+top_p = 1.0
+
+# ─── DeepSeek ────────────────────────────────────────────────────────
+# Source: https://api-docs.deepseek.com/quick_start/parameter_settings
+# Coding/Math: temp=0.0, General: temp=1.3, Creative: temp=1.5
+# NOTE: DeepSeek API applies nonlinear temp mapping. These are API values.
+# When running model directly: API temp 1.0 = model temp 0.3
+# We use temp=0.0 for eval (coding/math focus).
+
+[[model]]
+name = "DeepSeek V3.1"
+patterns = ["DeepSeek-V3.1"]
+reasoning = true
+temperature = 0.0
+
+# ─── GLM (ZhipuAI / THUDM) ──────────────────────────────────────────
+# Source: HuggingFace model cards + generation_config.json + docs.z.ai
+# GLM 4.5+: temp=1.0, top_p=0.95
+# Reasoning tasks: 131072 max_tokens; coding/SWE tasks: temp=0.7
+
+[[model]]
+name = "GLM-5"
+patterns = ["GLM-5"]
+reasoning = true
+temperature = 1.0
+top_p = 0.95
+max_tokens = 131072
+
+[[model]]
+name = "GLM 4.5 Air"
+patterns = ["GLM-4.5-Air"]
+reasoning = true
+temperature = 1.0
+top_p = 0.95
+
+[[model]]
+name = "GLM 4.7"
+patterns = ["GLM-4.7-"]
+reasoning = true
+temperature = 1.0
+top_p = 0.95
+max_tokens = 131072
+# Note: matches both GLM-4.7 and GLM-4.7-Flash
+
+# ─── Kimi (Moonshot AI) ─────────────────────────────────────────────
+# Source: HuggingFace model cards (moonshotai/Kimi-K2-*)
+# K2-Instruct: temp=0.6
+# K2-Thinking: temp=1.0, max_length=262144
+# K2.5: thinking temp=1.0, top_p=0.95; instant temp=0.6, top_p=0.95
+
+[[model]]
+name = "Kimi K2 Thinking"
+patterns = ["Kimi-K2-Thinking"]
+reasoning = true
+temperature = 1.0
+max_tokens = 131072
+
+[[model]]
+name = "Kimi K2.5"
+patterns = ["Kimi-K2.5"]
+reasoning = true
+temperature = 1.0
+top_p = 0.95
+max_tokens = 131072
+
+[[model]]
+name = "Kimi K2 Instruct"
+patterns = ["Kimi-K2-Instruct"]
+reasoning = false
+temperature = 0.6
+
+# ─── MiniMax ─────────────────────────────────────────────────────────
+# Source: HuggingFace model cards + generation_config.json
+# All models: temp=1.0, top_p=0.95, top_k=40
+
+[[model]]
+name = "MiniMax M2.5"
+patterns = ["MiniMax-M2.5"]
+reasoning = true
+temperature = 1.0
+top_p = 0.95
+
+[[model]]
+name = "MiniMax M2.1"
+patterns = ["MiniMax-M2.1"]
+reasoning = true
+temperature = 1.0
+top_p = 0.95
+
+# ─── Step (StepFun) ─────────────────────────────────────────────────
+# Source: HuggingFace model card (stepfun-ai/Step-3.5-Flash)
+# Reasoning: temp=1.0, top_p=0.95
+# General chat: temp=0.6, top_p=0.95
+# We use reasoning settings for eval.
+
+[[model]]
+name = "Step 3.5 Flash"
+patterns = ["Step-3.5-Flash"]
+reasoning = true
+temperature = 1.0
+top_p = 0.95
+
+# ─── Llama (Meta) ───────────────────────────────────────────────────
+# Source: generation_config.json + meta-llama/llama-models generation.py
+# All variants: temp=0.6, top_p=0.9
+
+[[model]]
+name = "Llama 3.2 1B"
+patterns = ["Llama-3.2-1B"]
+reasoning = false
+temperature = 0.6
+top_p = 0.9
+
+[[model]]
+name = "Llama 3.2 3B"
+patterns = ["Llama-3.2-3B"]
+reasoning = false
+temperature = 0.6
+top_p = 0.9
+
+[[model]]
+name = "Llama 3.1 8B"
+patterns = ["Llama-3.1-8B", "Meta-Llama-3.1-8B"]
+reasoning = false
+temperature = 0.6
+top_p = 0.9
+
+[[model]]
+name = "Llama 3.1 70B"
+patterns = ["Llama-3.1-70B", "Meta-Llama-3.1-70B"]
+reasoning = false
+temperature = 0.6
+top_p = 0.9
+
+[[model]]
+name = "Llama 3.3 70B"
+patterns = ["Llama-3.3-70B", "llama-3.3-70b"]
+reasoning = false
+temperature = 0.6
+top_p = 0.9

bench/exo_bench.py

@@ -24,6 +24,7 @@ import json
 import sys
 import time
 from collections.abc import Callable
+from concurrent.futures import ThreadPoolExecutor, as_completed
 from pathlib import Path
 from statistics import mean
 from typing import Any
@@ -252,6 +253,12 @@ def main() -> int:
     ap.add_argument(
         "--repeat", type=int, default=1, help="Repetitions per (pp,tg) pair."
     )
+    ap.add_argument(
+        "--concurrency",
+        nargs="+",
+        default=["1"],
+        help="Concurrency levels (ints). Accepts commas. E.g. --concurrency 1,2,4,8. Default 1.",
+    )
     ap.add_argument(
         "--warmup",
         type=int,
@@ -282,6 +289,10 @@ def main() -> int:
     if args.repeat <= 0:
         logger.error("--repeat must be >= 1")
         return 2
+    concurrency_list = parse_int_list(args.concurrency)
+    if not concurrency_list or any(c <= 0 for c in concurrency_list):
+        logger.error("--concurrency values must be >= 1")
+        return 2
 
     # Log pairing mode
     use_combinations = args.all_combinations or len(pp_list) != len(tg_list)
@@ -293,7 +304,9 @@ def main() -> int:
         logger.info(f"pp/tg mode: tandem (zip) - {len(pp_list)} pairs")
 
     client = ExoClient(args.host, args.port, timeout_s=args.timeout)
-    short_id, full_model_id = resolve_model_short_id(client, args.model)
+    short_id, full_model_id = resolve_model_short_id(
+        client, args.model, force_download=args.force_download
+    )
 
     tokenizer = load_tokenizer_for_bench(full_model_id)
     if tokenizer is None:
@@ -392,52 +405,131 @@ def main() -> int:
                 pp_tg_pairs = list(zip(pp_list, tg_list, strict=True))
 
             for pp, tg in pp_tg_pairs:
-                runs: list[dict[str, Any]] = []
-                for r in range(args.repeat):
-                    time.sleep(3)
-                    try:
-                        row, actual_pp_tokens = run_one_completion(
-                            client, full_model_id, pp, tg, prompt_sizer
+                for concurrency in concurrency_list:
+                    logger.info(f"--- pp={pp} tg={tg} concurrency={concurrency} ---")
+                    runs: list[dict[str, Any]] = []
+                    for r in range(args.repeat):
+                        time.sleep(3)
+
+                        if concurrency <= 1:
+                            # Sequential: single request
+                            try:
+                                row, actual_pp_tokens = run_one_completion(
+                                    client, full_model_id, pp, tg, prompt_sizer
+                                )
+                            except Exception as e:
+                                logger.error(e)
+                                continue
+                            row.update(
+                                {
+                                    "model_short_id": short_id,
+                                    "model_id": full_model_id,
+                                    "placement_sharding": sharding,
+                                    "placement_instance_meta": instance_meta,
+                                    "placement_nodes": n_nodes,
+                                    "instance_id": instance_id,
+                                    "pp_tokens": actual_pp_tokens,
+                                    "tg": tg,
+                                    "repeat_index": r,
+                                    "concurrency": 1,
+                                    **(
+                                        {"download_duration_s": download_duration_s}
+                                        if download_duration_s is not None
+                                        else {}
+                                    ),
+                                }
+                            )
+                            runs.append(row)
+                            all_rows.append(row)
+                        else:
+                            # Concurrent: fire N requests in parallel
+                            # Each thread gets its own ExoClient (separate HTTP connection)
+                            batch_results: list[tuple[dict[str, Any], int]] = []
+                            batch_errors = 0
+
+                            def _run_concurrent(
+                                idx: int, *, _pp: int = pp, _tg: int = tg
+                            ) -> tuple[dict[str, Any], int]:
+                                c = ExoClient(
+                                    args.host, args.port, timeout_s=args.timeout
+                                )
+                                return run_one_completion(
+                                    c, full_model_id, _pp, _tg, prompt_sizer
+                                )
+
+                            with ThreadPoolExecutor(max_workers=concurrency) as pool:
+                                futures = {
+                                    pool.submit(_run_concurrent, i): i
+                                    for i in range(concurrency)
+                                }
+                                for fut in as_completed(futures):
+                                    try:
+                                        batch_results.append(fut.result())
+                                    except Exception as e:
+                                        logger.error(f"Concurrent request failed: {e}")
+                                        batch_errors += 1
+
+                            for idx, (row, actual_pp_tokens) in enumerate(
+                                batch_results
+                            ):
+                                row.update(
+                                    {
+                                        "model_short_id": short_id,
+                                        "model_id": full_model_id,
+                                        "placement_sharding": sharding,
+                                        "placement_instance_meta": instance_meta,
+                                        "placement_nodes": n_nodes,
+                                        "instance_id": instance_id,
+                                        "pp_tokens": actual_pp_tokens,
+                                        "tg": tg,
+                                        "repeat_index": r,
+                                        "concurrency": concurrency,
+                                        "concurrent_index": idx,
+                                        **(
+                                            {"download_duration_s": download_duration_s}
+                                            if download_duration_s is not None
+                                            else {}
+                                        ),
+                                    }
+                                )
+                                runs.append(row)
+                                all_rows.append(row)
+
+                            if batch_results:
+                                valid_gen_tps = [
+                                    x["stats"]["generation_tps"]
+                                    for x, _ in batch_results
+                                    if x["stats"]["generation_tps"] > 0
+                                ]
+                                agg_gen_tps = (
+                                    mean(valid_gen_tps) if valid_gen_tps else 0.0
+                                )
+                                gen_tps = agg_gen_tps / concurrency
+                                logger.info(
+                                    f"[concurrent {concurrency}x]  "
+                                    f"agg_gen_tps={agg_gen_tps:.2f}  "
+                                    f"gen_tps={gen_tps:.2f}  "
+                                    f"errors={batch_errors}"
+                                )
+
+                    if runs:
+                        prompt_tps = mean(x["stats"]["prompt_tps"] for x in runs)
+                        gen_tps = mean(
+                            x["stats"]["generation_tps"] / x["concurrency"]
+                            for x in runs
+                        )
+                        ptok = mean(x["stats"]["prompt_tokens"] for x in runs)
+                        gtok = mean(x["stats"]["generation_tokens"] for x in runs)
+                        peak = mean(
+                            x["stats"]["peak_memory_usage"]["inBytes"] for x in runs
                         )
-                    except Exception as e:
-                        logger.error(e)
-                        continue
-                    row.update(
-                        {
-                            "model_short_id": short_id,
-                            "model_id": full_model_id,
-                            "placement_sharding": sharding,
-                            "placement_instance_meta": instance_meta,
-                            "placement_nodes": n_nodes,
-                            "instance_id": instance_id,
-                            "pp_tokens": actual_pp_tokens,
-                            "tg": tg,
-                            "repeat_index": r,
-                            **(
-                                {"download_duration_s": download_duration_s}
-                                if download_duration_s is not None
-                                else {}
-                            ),
-                        }
-                    )
-                    runs.append(row)
-                    all_rows.append(row)
 
-                if runs:
-                    prompt_tps = mean(x["stats"]["prompt_tps"] for x in runs)
-                    gen_tps = mean(x["stats"]["generation_tps"] for x in runs)
-                    ptok = mean(x["stats"]["prompt_tokens"] for x in runs)
-                    gtok = mean(x["stats"]["generation_tokens"] for x in runs)
-                    peak = mean(
-                        x["stats"]["peak_memory_usage"]["inBytes"] for x in runs
-                    )
-
-                    logger.info(
-                        f"prompt_tps={prompt_tps:.2f} gen_tps={gen_tps:.2f}    "
-                        f"prompt_tokens={ptok} gen_tokens={gtok}    "
-                        f"peak_memory={format_peak_memory(peak)}\n"
-                    )
-                time.sleep(2)
+                        logger.info(
+                            f"prompt_tps={prompt_tps:.2f} gen_tps={gen_tps:.2f}    "
+                            f"prompt_tokens={ptok} gen_tokens={gtok}    "
+                            f"peak_memory={format_peak_memory(peak)}\n"
+                        )
+                    time.sleep(2)
         finally:
             try:
                 client.request_json("DELETE", f"/instance/{instance_id}")

bench/harness.py

@@ -165,7 +165,9 @@ def wait_for_instance_gone(
     raise TimeoutError(f"Instance {instance_id} did not get deleted within {timeout=}")
 
 
-def resolve_model_short_id(client: ExoClient, model_arg: str) -> tuple[str, str]:
+def resolve_model_short_id(
+    client: ExoClient, model_arg: str, *, force_download: bool = False
+) -> tuple[str, str]:
     models = client.request_json("GET", "/models") or {}
     data = models.get("data") or []
 
@@ -181,6 +183,16 @@ def resolve_model_short_id(client: ExoClient, model_arg: str) -> tuple[str, str]
             full_id = str(m["hugging_face_id"])
             return short_id, full_id
 
+    if force_download and "/" in model_arg:
+        logger.info(f"Model not in /models, adding from HuggingFace: {model_arg}")
+        result = client.request_json(
+            "POST", "/models/add", body={"model_id": model_arg}
+        )
+        if result:
+            short_id = str(result.get("name") or model_arg.rsplit("/", 1)[-1])
+            full_id = str(result.get("hugging_face_id") or model_arg)
+            return short_id, full_id
+
     raise ValueError(f"Model not found in /models: {model_arg}")
 
 
@@ -445,6 +457,11 @@ def add_common_instance_args(ap: argparse.ArgumentParser) -> None:
         "--port", type=int, default=int(os.environ.get("EXO_PORT", "52415"))
     )
     ap.add_argument("--model", required=True, help="Model short id or huggingface id")
+    ap.add_argument(
+        "--force-download",
+        action="store_true",
+        help="If model not in /models, add it from HuggingFace via exo and download.",
+    )
     ap.add_argument(
         "--max-nodes",
         type=int,

bench/parallel_requests.py

@@ -0,0 +1,255 @@
+# type: ignore
+import argparse
+import asyncio
+import sys
+import termios
+import time
+import tty
+
+import aiohttp
+
+NUM_REQUESTS = 10
+BASE_URL = ""
+
+QUESTIONS = [
+    "What is the capital of Australia?",
+    "How many bones are in the human body?",
+    "What year did World War II end?",
+    "What is the speed of light in meters per second?",
+    "Who wrote Romeo and Juliet?",
+    "What is the chemical formula for water?",
+    "How many planets are in our solar system?",
+    "What is the largest ocean on Earth?",
+    "Who painted the Mona Lisa?",
+    "What is the boiling point of water in Celsius?",
+]
+
+
+def write(s: str) -> None:
+    sys.stdout.write(s)
+
+
+# ---------------------------------------------------------------------------
+# Model picker (same style as exo_eval)
+# ---------------------------------------------------------------------------
+
+
+def fetch_models() -> list[str]:
+    import json
+    import urllib.request
+
+    with urllib.request.urlopen(f"{BASE_URL}/state") as resp:
+        data = json.loads(resp.read())
+    model_ids: set[str] = set()
+    for instance in data.get("instances", {}).values():
+        for variant in instance.values():
+            sa = variant.get("shardAssignments", {})
+            model_id = sa.get("modelId")
+            if model_id:
+                model_ids.add(model_id)
+    return sorted(model_ids)
+
+
+def pick_model() -> str | None:
+    models = fetch_models()
+    if not models:
+        print("No models found.")
+        return None
+
+    cursor = 0
+    total_lines = len(models) + 4
+
+    def render(first: bool = False) -> None:
+        if not first:
+            write(f"\033[{total_lines}A")
+        write("\033[J")
+        write("\033[1mSelect model\033[0m (up/down, enter confirm, q quit)\r\n\r\n")
+        for i, model in enumerate(models):
+            line = f"  {'>' if i == cursor else ' '} {model}"
+            write(f"\033[7m{line}\033[0m\r\n" if i == cursor else f"{line}\r\n")
+        write("\r\n")
+        sys.stdout.flush()
+
+    fd = sys.stdin.fileno()
+    old = termios.tcgetattr(fd)
+    try:
+        tty.setraw(fd)
+        write("\033[?25l")
+        render(first=True)
+        while True:
+            ch = sys.stdin.read(1)
+            if ch in ("q", "\x03"):
+                write("\033[?25h\033[0m\r\n")
+                return None
+            elif ch in ("\r", "\n"):
+                break
+            elif ch == "\x1b":
+                seq = sys.stdin.read(2)
+                if seq == "[A":
+                    cursor = (cursor - 1) % len(models)
+                elif seq == "[B":
+                    cursor = (cursor + 1) % len(models)
+            render()
+    finally:
+        termios.tcsetattr(fd, termios.TCSADRAIN, old)
+        write(f"\033[{total_lines}A\033[J")  # clear picker UI
+        write("\033[?25h\033[0m")
+        sys.stdout.flush()
+
+    return models[cursor]
+
+
+# ---------------------------------------------------------------------------
+# Parallel requests
+# ---------------------------------------------------------------------------
+
+statuses: list[str] = []
+times: list[str] = []
+previews: list[str] = []
+tokens: list[str] = []
+full_responses: list[dict | None] = []
+total_lines = 0
+start_time: float = 0
+selected_model: str = ""
+
+
+def render_progress(first: bool = False) -> None:
+    if not first:
+        write(f"\033[{total_lines}A")
+    write("\033[J")
+    elapsed = time.monotonic() - start_time if start_time else 0
+    done = sum(1 for s in statuses if s == "done")
+    write(
+        f"\033[1m{selected_model}\033[0m  [{done}/{NUM_REQUESTS}]  {elapsed:.1f}s\r\n\r\n"
+    )
+
+    for i in range(NUM_REQUESTS):
+        q = QUESTIONS[i % len(QUESTIONS)]
+        status = statuses[i]
+        if status == "pending":
+            color = "\033[33m"  # yellow
+        elif status == "running":
+            color = "\033[36m"  # cyan
+        elif status == "done":
+            color = "\033[32m"  # green
+        else:
+            color = "\033[31m"  # red
+        write(
+            f"  {i:>2}  {color}{status:<8}\033[0m  {times[i]:>6}  {tokens[i]:>5}tok  {q[:40]:<40}  {previews[i][:50]}\r\n"
+        )
+
+    write("\r\n")
+    sys.stdout.flush()
+
+
+async def send_request(
+    session: aiohttp.ClientSession, i: int, lock: asyncio.Lock
+) -> None:
+    payload = {
+        "model": selected_model,
+        "messages": [{"role": "user", "content": QUESTIONS[i % len(QUESTIONS)]}],
+        "max_tokens": 1024,
+    }
+    statuses[i] = "running"
+    async with lock:
+        render_progress()
+    t0 = time.monotonic()
+    try:
+        async with session.post(
+            f"{BASE_URL}/v1/chat/completions", json=payload
+        ) as resp:
+            data = await resp.json()
+            elapsed = time.monotonic() - t0
+            full_responses[i] = data
+            times[i] = f"{elapsed:.1f}s"
+            if resp.status == 200:
+                choice = data["choices"][0]
+                msg = choice["message"]
+                content = msg.get("content", "")
+                previews[i] = content[:50].replace("\n", " ") or "(empty)"
+                if "usage" in data:
+                    tokens[i] = str(data["usage"].get("total_tokens", ""))
+                statuses[i] = "done"
+            else:
+                statuses[i] = f"err:{resp.status}"
+                previews[i] = str(data.get("error", {}).get("message", ""))[:50]
+    except Exception as e:
+        elapsed = time.monotonic() - t0
+        times[i] = f"{elapsed:.1f}s"
+        statuses[i] = "error"
+        previews[i] = str(e)[:50]
+    async with lock:
+        render_progress()
+
+
+async def run_requests(print_stdout: bool = False) -> None:
+    global start_time, total_lines, statuses, times, previews, tokens, full_responses
+
+    statuses = ["pending"] * NUM_REQUESTS
+    times = ["-"] * NUM_REQUESTS
+    previews = ["-"] * NUM_REQUESTS
+    tokens = ["-"] * NUM_REQUESTS
+    full_responses = [None] * NUM_REQUESTS
+    total_lines = NUM_REQUESTS + 4
+
+    write("\033[?25l")  # hide cursor
+    start_time = time.monotonic()
+    render_progress(first=True)
+    lock = asyncio.Lock()
+    try:
+        async with aiohttp.ClientSession() as session:
+            tasks = [send_request(session, i, lock) for i in range(NUM_REQUESTS)]
+            await asyncio.gather(*tasks)
+        total = time.monotonic() - start_time
+        write(
+            f"\033[1m=== All {NUM_REQUESTS} requests done in {total:.1f}s ===\033[0m\r\n\r\n"
+        )
+
+        if print_stdout:
+            for i in range(NUM_REQUESTS):
+                data = full_responses[i]
+                if not data or "choices" not in data:
+                    continue
+                choice = data["choices"][0]
+                msg = choice["message"]
+                q = QUESTIONS[i % len(QUESTIONS)]
+                write(f"\033[1m--- #{i}: {q} ---\033[0m\r\n")
+                if msg.get("reasoning_content"):
+                    write(f"\033[2m[Thinking]: {msg['reasoning_content']}\033[0m\r\n")
+                write(f"{msg.get('content', '')}\r\n")
+                if "usage" in data:
+                    u = data["usage"]
+                    write(
+                        f"\033[2m[Usage: prompt={u.get('prompt_tokens')}, "
+                        f"completion={u.get('completion_tokens')}, "
+                        f"total={u.get('total_tokens')}]\033[0m\r\n"
+                    )
+                write("\r\n")
+    finally:
+        write("\033[?25h")  # show cursor
+        sys.stdout.flush()
+
+
+def main() -> None:
+    global selected_model, BASE_URL
+    parser = argparse.ArgumentParser(
+        description="Send parallel requests to an exo cluster"
+    )
+    parser.add_argument(
+        "--host", required=True, help="Hostname of the exo node (e.g. s1)"
+    )
+    parser.add_argument("--port", type=int, default=52415, help="Port (default: 52415)")
+    parser.add_argument(
+        "--stdout", action="store_true", help="Print full responses after completion"
+    )
+    args = parser.parse_args()
+    BASE_URL = f"http://{args.host}:{args.port}"
+    model = pick_model()
+    if not model:
+        return
+    selected_model = model
+    asyncio.run(run_requests(print_stdout=args.stdout))
+
+
+if __name__ == "__main__":
+    main()

bench/pyproject.toml

@@ -4,13 +4,18 @@ version = "0.1.0"
 description = "Benchmarking tool for exo distributed inference"
 requires-python = ">=3.13"
 dependencies = [
-    "httpx>=0.27.0",
-    "loguru>=0.7.3",
-    "transformers>=5.0.0",
-    "huggingface-hub>=0.33.4",
-    "tiktoken>=0.12.0",
-    "jinja2>=3.1.0",
-    "protobuf>=5.29.0",
+  "httpx>=0.27.0",
+  "loguru>=0.7.3",
+  "transformers>=5.0.0",
+  "huggingface-hub>=0.33.4",
+  "tiktoken>=0.12.0",
+  "jinja2>=3.1.0",
+  "protobuf>=5.29.0",
+  "datasets>=2.0.0",
+  "math-verify>=0.7.0",
+  "lm-eval[api,math]>=0.4.0",
+  "human-eval>=1.0.3",
+  "numpy>=1.24.0",
 ]
 
 [build-system]

bench/single-m3-ultra.toml

@@ -2,10 +2,10 @@
 #
 # Shared constraints applied to ALL benchmarks in this file.
 constraints = [
-    "All(MacOsBuild(=25D125))",
-    "Hosts(=1)",
-    "All(Chip(m3_ultra))",
-    "All(GpuCores(=80))",
+  "All(MacOsBuild(=25D125))",
+  "Hosts(=1)",
+  "All(Chip(m3_ultra))",
+  "All(GpuCores(=80))",
 ]
 
 [topology]

bench/vendor/lcb_testing_util.py

@@ -0,0 +1,593 @@
+# type: ignore
+# Vendored from LiveCodeBench (https://github.com/LiveCodeBench/LiveCodeBench)
+# File: lcb_runner/evaluation/testing_util.py
+# License: MIT
+# Vendored 2026-03-07 — do not modify without updating from upstream.
+
+import ast
+import faulthandler
+import json
+import platform
+
+# to run the solution files we're using a timing based approach
+import signal
+import sys
+import time
+
+# used for debugging to time steps
+from datetime import datetime
+from decimal import Decimal
+from enum import Enum
+from io import StringIO
+
+# from pyext import RuntimeModule
+from types import ModuleType
+
+# used for testing the code that reads from input
+from unittest.mock import mock_open, patch
+
+import_string = "from string import *\nfrom re import *\nfrom datetime import *\nfrom collections import *\nfrom heapq import *\nfrom bisect import *\nfrom copy import *\nfrom math import *\nfrom random import *\nfrom statistics import *\nfrom itertools import *\nfrom functools import *\nfrom operator import *\nfrom io import *\nfrom sys import *\nfrom json import *\nfrom builtins import *\nfrom typing import *\nimport string\nimport re\nimport datetime\nimport collections\nimport heapq\nimport bisect\nimport copy\nimport math\nimport random\nimport statistics\nimport itertools\nimport functools\nimport operator\nimport io\nimport sys\nimport json\nsys.setrecursionlimit(50000)\n"
+
+
+def truncatefn(s, length=300):
+    if isinstance(s, str):
+        pass
+    else:
+        s = str(s)
+    if len(s) <= length:
+        return s
+
+    return s[: length // 2] + "...(truncated) ..." + s[-length // 2 :]
+
+
+class CODE_TYPE(Enum):
+    call_based = 0
+    standard_input = 1
+
+
+# stuff for setting up signal timer
+class TimeoutException(Exception):
+    pass
+
+
+def timeout_handler(signum, frame):
+    print("timeout occured: alarm went off")
+    raise TimeoutException
+
+
+# used to capture stdout as a list
+# from https://stackoverflow.com/a/16571630/6416660
+# alternative use redirect_stdout() from contextlib
+class Capturing(list):
+    def __enter__(self):
+        self._stdout = sys.stdout
+        sys.stdout = self._stringio = StringIO()
+        # Make closing the StringIO a no-op
+        self._stringio.close = lambda x: 1
+        return self
+
+    def __exit__(self, *args):
+        self.append(self._stringio.getvalue())
+        del self._stringio  # free up some memory
+        sys.stdout = self._stdout
+
+
+# Custom mock for sys.stdin that supports buffer attribute
+class MockStdinWithBuffer:
+    def __init__(self, inputs: str):
+        self.inputs = inputs
+        self._stringio = StringIO(inputs)
+        self.buffer = MockBuffer(inputs)
+
+    def read(self, *args):
+        return self.inputs
+
+    def readline(self, *args):
+        return self._stringio.readline(*args)
+
+    def readlines(self, *args):
+        return self.inputs.split("\n")
+
+    def __getattr__(self, name):
+        # Delegate other attributes to StringIO
+        return getattr(self._stringio, name)
+
+
+class MockBuffer:
+    def __init__(self, inputs: str):
+        self.inputs = inputs.encode("utf-8")  # Convert to bytes
+
+    def read(self, *args):
+        # Return as byte strings that can be split
+        return self.inputs
+
+    def readline(self, *args):
+        return self.inputs.split(b"\n")[0] + b"\n"
+
+
+def clean_if_name(code: str) -> str:
+    try:
+        astree = ast.parse(code)
+        last_block = astree.body[-1]
+        if isinstance(last_block, ast.If):
+            condition = last_block.test
+            if ast.unparse(condition).strip() == "__name__ == '__main__'":
+                code = (
+                    ast.unparse(astree.body[:-1]) + "\n" + ast.unparse(last_block.body)  # type: ignore
+                )
+    except:
+        pass
+
+    return code
+
+
+def make_function(code: str) -> str:
+    try:
+        import_stmts = []
+        all_other_stmts = []
+        astree = ast.parse(code)
+        for stmt in astree.body:
+            if isinstance(stmt, (ast.Import, ast.ImportFrom)):
+                import_stmts.append(stmt)
+            else:
+                all_other_stmts.append(stmt)
+
+        function_ast = ast.FunctionDef(
+            name="wrapped_function",
+            args=ast.arguments(
+                posonlyargs=[], args=[], kwonlyargs=[], kw_defaults=[], defaults=[]
+            ),
+            body=all_other_stmts,
+            decorator_list=[],
+            lineno=-1,
+        )
+        main_code = (
+            import_string
+            + "\n"
+            + ast.unparse(import_stmts)
+            + "\n"
+            + ast.unparse(function_ast)
+        )
+        return main_code
+    except Exception:
+        return code
+
+
+def call_method(method, inputs):
+    if isinstance(inputs, list):
+        inputs = "\n".join(inputs)
+
+    inputs_line_iterator = iter(inputs.split("\n"))
+
+    # Create custom stdin mock with buffer support
+    mock_stdin = MockStdinWithBuffer(inputs)
+
+    # sys.setrecursionlimit(10000)
+
+    # @patch('builtins.input', side_effect=inputs.split("\n"))
+    @patch("builtins.open", mock_open(read_data=inputs))
+    @patch("sys.stdin", mock_stdin)  # Use our custom mock instead of StringIO
+    @patch("sys.stdin.readline", lambda *args: next(inputs_line_iterator))
+    @patch("sys.stdin.readlines", lambda *args: inputs.split("\n"))
+    @patch("sys.stdin.read", lambda *args: inputs)
+    # @patch('sys.stdout.write', print)
+    def _inner_call_method(_method):
+        try:
+            return _method()
+        except SystemExit:
+            pass
+        finally:
+            pass
+
+    return _inner_call_method(method)
+
+
+def get_function(compiled_sol, fn_name: str):  # type: ignore
+    try:
+        assert hasattr(compiled_sol, fn_name)
+        return getattr(compiled_sol, fn_name)
+    except Exception:
+        return
+
+
+def compile_code(code: str, timeout: int):
+    signal.alarm(timeout)
+    try:
+        tmp_sol = ModuleType("tmp_sol", "")
+        exec(code, tmp_sol.__dict__)
+        if "class Solution" in code:
+            # leetcode wraps solutions in `Solution`
+            # this is a hack to check if it is leetcode solution or not
+            # currently livecodebench only supports LeetCode but
+            # else condition allows future extensibility to other platforms
+            compiled_sol = tmp_sol.Solution()
+        else:
+            # do nothing in the other case since function is accesible
+            compiled_sol = tmp_sol
+
+        assert compiled_sol is not None
+    finally:
+        signal.alarm(0)
+
+    return compiled_sol
+
+
+def convert_line_to_decimals(line: str) -> tuple[bool, list[Decimal]]:
+    try:
+        decimal_line = [Decimal(elem) for elem in line.split()]
+    except:
+        return False, []
+    return True, decimal_line
+
+
+def get_stripped_lines(val: str):
+    ## you don't want empty lines to add empty list after splitlines!
+    val = val.strip()
+
+    return [val_line.strip() for val_line in val.split("\n")]
+
+
+def grade_call_based(
+    code: str, all_inputs: list, all_outputs: list, fn_name: str, timeout: int
+):
+    # call-based clean up logic
+    # need to wrap in try-catch logic after to catch the correct errors, but for now this is fine.
+    code = import_string + "\n\n" + code
+    compiled_sol = compile_code(code, timeout)
+
+    if compiled_sol is None:
+        return
+
+    method = get_function(compiled_sol, fn_name)
+
+    if method is None:
+        return
+
+    all_inputs = [
+        [json.loads(line) for line in inputs.split("\n")] for inputs in all_inputs
+    ]
+
+    all_outputs = [json.loads(output) for output in all_outputs]
+
+    total_execution = 0
+    all_results = []
+    for idx, (gt_inp, gt_out) in enumerate(zip(all_inputs, all_outputs)):
+        signal.alarm(timeout)
+        faulthandler.enable()
+        try:
+            # can lock here so time is useful
+            start = time.time()
+            prediction = method(*gt_inp)
+            total_execution += time.time() - start
+            signal.alarm(0)
+
+            # don't penalize model if it produces tuples instead of lists
+            # ground truth sequences are not tuples
+            if isinstance(prediction, tuple):
+                prediction = list(prediction)
+
+            tmp_result = prediction == gt_out
+
+            # handle floating point comparisons
+
+            all_results.append(tmp_result)
+
+            if not tmp_result:
+                return all_results, {
+                    "output": truncatefn(prediction),
+                    "inputs": truncatefn(gt_inp),
+                    "expected": truncatefn(gt_out),
+                    "error_code": -2,
+                    "error_message": "Wrong Answer",
+                }
+        except Exception as e:
+            signal.alarm(0)
+            if "timeoutexception" in repr(e).lower():
+                all_results.append(-3)
+                return all_results, {
+                    "error": repr(e),
+                    "error_code": -3,
+                    "error_message": "Time Limit Exceeded",
+                    "inputs": truncatefn(gt_inp),
+                    "expected": truncatefn(gt_out),
+                }
+            else:
+                all_results.append(-4)
+                return all_results, {
+                    "error": repr(e),
+                    "error_code": -4,
+                    "error_message": "Runtime Error",
+                    "inputs": truncatefn(gt_inp),
+                    "expected": truncatefn(gt_out),
+                }
+
+        finally:
+            signal.alarm(0)
+            faulthandler.disable()
+
+    return all_results, {"execution time": total_execution}
+
+
+def grade_stdio(
+    code: str,
+    all_inputs: list,
+    all_outputs: list,
+    timeout: int,
+):
+    ## runtime doesn't interact well with __name__ == '__main__'
+    code = clean_if_name(code)
+
+    ## we wrap the given code inside another function
+    code = make_function(code)
+
+    compiled_sol = compile_code(code, timeout)
+    if compiled_sol is None:
+        return
+
+    method = get_function(compiled_sol, "wrapped_function")
+
+    if method is None:
+        return
+
+    all_results = []
+    total_execution_time = 0
+    for idx, (gt_inp, gt_out) in enumerate(zip(all_inputs, all_outputs)):
+        signal.alarm(timeout)
+        faulthandler.enable()
+
+        signal.alarm(timeout)
+        with Capturing() as captured_output:
+            try:
+                start = time.time()
+                call_method(method, gt_inp)
+                total_execution_time += time.time() - start
+                # reset the alarm
+                signal.alarm(0)
+            except Exception as e:
+                signal.alarm(0)
+                if "timeoutexception" in repr(e).lower():
+                    all_results.append(-3)
+                    return all_results, {
+                        "error": repr(e),
+                        "error_code": -3,
+                        "error_message": "Time Limit Exceeded",
+                        "inputs": truncatefn(gt_inp),
+                        "expected": truncatefn(gt_out),
+                    }
+                else:
+                    all_results.append(-4)
+                    return all_results, {
+                        "error": repr(e),
+                        "error_code": -4,
+                        "error_message": "Runtime Error",
+                        "inputs": truncatefn(gt_inp),
+                        "expected": truncatefn(gt_out),
+                    }
+
+            finally:
+                signal.alarm(0)
+                faulthandler.disable()
+
+        prediction = captured_output[0]
+
+        stripped_prediction_lines = get_stripped_lines(prediction)
+        stripped_gt_out_lines = get_stripped_lines(gt_out)
+
+        ## WA happens in multiple circumstances
+        ## so cache the return to make it clean!
+        WA_send_args = {
+            "output": truncatefn(prediction),
+            "inputs": truncatefn(gt_inp),
+            "expected": truncatefn(gt_out),
+            "error_code": -2,
+        }
+
+        if len(stripped_prediction_lines) != len(stripped_gt_out_lines):
+            all_results.append(-2)
+            WA_send_args["error_message"] = "Wrong answer: mismatched output length"
+            return all_results, WA_send_args
+
+        for output_line_idx, (
+            stripped_prediction_line,
+            stripped_gt_out_line,
+        ) in enumerate(zip(stripped_prediction_lines, stripped_gt_out_lines)):
+            WA_send_args["error_message"] = (
+                f"Wrong answer at {output_line_idx=}: {truncatefn(stripped_prediction_line)} != {truncatefn(stripped_gt_out_line)}"
+            )
+
+            ## CASE 1: exact match
+            if stripped_prediction_line == stripped_gt_out_line:
+                continue
+
+            ## CASE 2: element-wise comparision
+            ## if there are floating elements
+            ## use `decimal` library for good floating point comparision
+            ## otherwise gotcha: np.isclose(50000000000000000, 50000000000000001) = True
+            ## note that we should always be able to convert to decimals
+
+            success, decimal_prediction_line = convert_line_to_decimals(
+                stripped_prediction_line
+            )
+            if not success:
+                all_results.append(-2)
+                return all_results, WA_send_args
+            success, decimal_gtout_line = convert_line_to_decimals(stripped_gt_out_line)
+            if not success:
+                all_results.append(-2)
+                return all_results, WA_send_args
+
+            if decimal_prediction_line == decimal_gtout_line:
+                continue
+
+            all_results.append(-2)
+            return all_results, WA_send_args
+        all_results.append(True)
+
+    return all_results, {"execution time": total_execution_time}
+
+
+def run_test(sample, test=None, debug=False, timeout=6):
+    """
+    if test(generated_code) is not None it'll try to run the code.
+    otherwise it'll just return an input and output pair.
+    """
+    signal.signal(signal.SIGALRM, timeout_handler)
+
+    # Disable functionalities that can make destructive changes to the test.
+    # max memory is set to 4GB
+    reliability_guard()
+
+    if debug:
+        print(f"start = {datetime.now().time()}")
+
+    try:
+        in_outs = json.loads(sample["input_output"])
+    except ValueError as e:
+        raise e
+        in_outs = None
+
+    if in_outs:
+        if in_outs.get("fn_name") is None:
+            which_type = CODE_TYPE.standard_input  # Standard input
+            method_name = None
+
+        else:
+            which_type = CODE_TYPE.call_based  # Call-based
+            method_name = in_outs["fn_name"]
+
+    if debug:
+        print(f"loaded input_output = {datetime.now().time()}")
+
+    if test is None:
+        assert False, "should not happen: test code is none"
+        return in_outs, {"error": "no test code provided"}
+    elif test is not None:
+        results = []
+        sol = import_string
+        if debug:
+            print(f"loading test code = {datetime.now().time()}")
+
+        if which_type == CODE_TYPE.call_based:
+            signal.alarm(timeout)
+            try:
+                results, metadata = grade_call_based(
+                    code=test,
+                    all_inputs=in_outs["inputs"],
+                    all_outputs=in_outs["outputs"],
+                    fn_name=method_name,
+                    timeout=timeout,
+                )
+                return results, metadata
+            except Exception as e:
+                return [-4], {
+                    "error_code": -4,
+                    "error_message": f"Error during testing: {e}",
+                }
+            finally:
+                signal.alarm(0)
+        elif which_type == CODE_TYPE.standard_input:
+            # sol
+            # if code has if __name__ == "__main__": then remove it
+
+            signal.alarm(timeout)
+            try:
+                results, metadata = grade_stdio(
+                    code=test,
+                    all_inputs=in_outs["inputs"],
+                    all_outputs=in_outs["outputs"],
+                    timeout=timeout,
+                )
+                return results, metadata
+            except Exception as e:
+                return [-4], {
+                    "error_code": -4,
+                    "error_message": f"Error during testing: {e}",
+                }
+            finally:
+                signal.alarm(0)
+
+
+def reliability_guard(maximum_memory_bytes=None):
+    """
+    This disables various destructive functions and prevents the generated code
+    from interfering with the test (e.g. fork bomb, killing other processes,
+    removing filesystem files, etc.)
+    WARNING
+    This function is NOT a security sandbox. Untrusted code, including, model-
+    generated code, should not be blindly executed outside of one. See the
+    Codex paper for more information about OpenAI's code sandbox, and proceed
+    with caution.
+    """
+
+    if maximum_memory_bytes is not None:
+        import resource
+
+        resource.setrlimit(
+            resource.RLIMIT_AS, (maximum_memory_bytes, maximum_memory_bytes)
+        )
+        resource.setrlimit(
+            resource.RLIMIT_DATA, (maximum_memory_bytes, maximum_memory_bytes)
+        )
+        if not platform.uname().system == "Darwin":
+            resource.setrlimit(
+                resource.RLIMIT_STACK, (maximum_memory_bytes, maximum_memory_bytes)
+            )
+
+    faulthandler.disable()
+
+    import builtins
+
+    # builtins.exit = None
+    builtins.quit = None
+
+    import os
+
+    os.environ["OMP_NUM_THREADS"] = "1"
+
+    os.kill = None
+    os.system = None
+    os.putenv = None
+    os.remove = None
+    os.removedirs = None
+    os.rmdir = None
+    os.fchdir = None
+    os.setuid = None
+    os.fork = None
+    os.forkpty = None
+    os.killpg = None
+    os.rename = None
+    os.renames = None
+    os.truncate = None
+    os.replace = None
+    os.unlink = None
+    os.fchmod = None
+    os.fchown = None
+    os.chmod = None
+    os.chown = None
+    os.chroot = None
+    os.fchdir = None
+    os.lchflags = None
+    os.lchmod = None
+    os.lchown = None
+    os.getcwd = None
+    os.chdir = None
+
+    import shutil
+
+    shutil.rmtree = None
+    shutil.move = None
+    shutil.chown = None
+
+    import subprocess
+
+    subprocess.Popen = None
+
+    __builtins__["help"] = None
+
+    import sys
+
+    sys.modules["ipdb"] = None
+    sys.modules["joblib"] = None
+    sys.modules["resource"] = None
+    sys.modules["psutil"] = None
+    sys.modules["tkinter"] = None

dashboard/src/lib/components/ChatForm.svelte

@@ -1,9 +1,6 @@
 <script lang="ts">
   import {
     isLoading,
-    sendMessage,
-    generateImage,
-    editImage,
     editingImage,
     clearEditingImage,
     selectedChatModel,
@@ -28,7 +25,7 @@
     modelTasks?: Record<string, string[]>;
     modelCapabilities?: Record<string, string[]>;
     onSend?: () => void;
-    onAutoSend?: (
+    onAutoSend: (
       content: string,
       files?: {
         id: string;
@@ -216,49 +213,10 @@
     uploadedFiles = [];
     resetTextareaHeight();
 
-    // When onAutoSend is provided, the parent controls all send logic
-    // (including launching non-running models before sending)
-    if (onAutoSend) {
-      onAutoSend(content, files);
-      onSend?.();
-      setTimeout(() => textareaRef?.focus(), 10);
-      return;
-    }
-
-    // Use image editing if in edit mode
-    if (isEditMode && currentEditingImage && content) {
-      editImage(content, currentEditingImage.imageDataUrl);
-    }
-    // If user attached an image with an ImageToImage model, use edit endpoint
-    else if (
-      currentModel &&
-      modelSupportsImageEditing(currentModel) &&
-      files.length > 0 &&
-      content
-    ) {
-      // Use the first attached image for editing
-      const imageFile = files[0];
-      if (imageFile.preview) {
-        editImage(content, imageFile.preview);
-      }
-    } else if (
-      currentModel &&
-      modelSupportsTextToImage(currentModel) &&
-      content
-    ) {
-      // Use image generation for text-to-image models
-      generateImage(content);
-    } else {
-      sendMessage(
-        content,
-        files,
-        modelSupportsThinking() ? thinkingEnabled : null,
-      );
-    }
-
+    // Parent controls all send logic (including image routing,
+    // launching non-running models before sending, etc.)
+    onAutoSend(content, files);
     onSend?.();
-
-    // Refocus the textarea after sending
     setTimeout(() => textareaRef?.focus(), 10);
   }
 

dashboard/src/lib/components/ChatSidebar.svelte

@@ -18,12 +18,18 @@
     class?: string;
     onNewChat?: () => void;
     onSelectConversation?: () => void;
+    isMobileDrawer?: boolean;
+    isOpen?: boolean;
+    onClose?: () => void;
   }
 
   let {
     class: className = "",
     onNewChat,
     onSelectConversation,
+    isMobileDrawer = false,
+    isOpen = false,
+    onClose,
   }: Props = $props();
 
   const conversationList = $derived(conversations());
@@ -53,6 +59,10 @@
   function handleSelectConversation(id: string) {
     onSelectConversation?.();
     loadConversation(id);
+    // Close mobile drawer when selecting a conversation
+    if (isMobileDrawer && isOpen) {
+      onClose?.();
+    }
   }
 
   function handleStartEdit(id: string, name: string, event: MouseEvent) {
@@ -252,9 +262,7 @@
   }
 </script>
 
-<aside
-  class="flex flex-col h-full bg-exo-dark-gray border-r border-exo-yellow/10 {className}"
->
+{#snippet sidebarContent()}
   <!-- Header -->
   <div class="p-4">
     <button
@@ -591,4 +599,30 @@
       </button>
     </div>
   </div>
-</aside>
+{/snippet}
+
+{#if isMobileDrawer}
+  <!-- Mobile drawer with overlay -->
+  {#if isOpen}
+    <!-- Overlay backdrop -->
+    <button
+      type="button"
+      class="fixed inset-0 bg-black/60 backdrop-blur-sm z-40 md:hidden"
+      onclick={() => onClose?.()}
+      aria-label="Close sidebar"
+    ></button>
+    <!-- Drawer panel -->
+    <aside
+      class="fixed left-0 top-0 bottom-0 w-72 bg-exo-dark-gray border-r border-exo-yellow/10 z-50 flex flex-col md:hidden"
+    >
+      {@render sidebarContent()}
+    </aside>
+  {/if}
+{:else}
+  <!-- Desktop sidebar -->
+  <aside
+    class="flex flex-col h-full bg-exo-dark-gray border-r border-exo-yellow/10 {className}"
+  >
+    {@render sidebarContent()}
+  </aside>
+{/if}

dashboard/src/lib/components/FamilyLogos.svelte

@@ -82,6 +82,12 @@
       d="M12.025 1.13c-5.77 0-10.449 4.647-10.449 10.378 0 1.112.178 2.181.503 3.185.064-.222.203-.444.416-.577a.96.96 0 0 1 .524-.15c.293 0 .584.124.84.284.278.173.48.408.71.694.226.282.458.611.684.951v-.014c.017-.324.106-.622.264-.874s.403-.487.762-.543c.3-.047.596.06.787.203s.31.313.4.467c.15.257.212.468.233.542.01.026.653 1.552 1.657 2.54.616.605 1.01 1.223 1.082 1.912.055.537-.096 1.059-.38 1.572.637.121 1.294.187 1.967.187.657 0 1.298-.063 1.921-.178-.287-.517-.44-1.041-.384-1.581.07-.69.465-1.307 1.081-1.913 1.004-.987 1.647-2.513 1.657-2.539.021-.074.083-.285.233-.542.09-.154.208-.323.4-.467a1.08 1.08 0 0 1 .787-.203c.359.056.604.29.762.543s.247.55.265.874v.015c.225-.34.457-.67.683-.952.23-.286.432-.52.71-.694.257-.16.547-.284.84-.285a.97.97 0 0 1 .524.151c.228.143.373.388.43.625l.006.04a10.3 10.3 0 0 0 .534-3.273c0-5.731-4.678-10.378-10.449-10.378M8.327 6.583a1.5 1.5 0 0 1 .713.174 1.487 1.487 0 0 1 .617 2.013c-.183.343-.762-.214-1.102-.094-.38.134-.532.914-.917.71a1.487 1.487 0 0 1 .69-2.803m7.486 0a1.487 1.487 0 0 1 .689 2.803c-.385.204-.536-.576-.916-.71-.34-.12-.92.437-1.103.094a1.487 1.487 0 0 1 .617-2.013 1.5 1.5 0 0 1 .713-.174m-10.68 1.55a.96.96 0 1 1 0 1.921.96.96 0 0 1 0-1.92m13.838 0a.96.96 0 1 1 0 1.92.96.96 0 0 1 0-1.92M8.489 11.458c.588.01 1.965 1.157 3.572 1.164 1.607-.007 2.984-1.155 3.572-1.164.196-.003.305.12.305.454 0 .886-.424 2.328-1.563 3.202-.22-.756-1.396-1.366-1.63-1.32q-.011.001-.02.006l-.044.026-.01.008-.03.024q-.018.017-.035.036l-.032.04a1 1 0 0 0-.058.09l-.014.025q-.049.088-.11.19a1 1 0 0 1-.083.116 1.2 1.2 0 0 1-.173.18q-.035.029-.075.058a1.3 1.3 0 0 1-.251-.243 1 1 0 0 1-.076-.107c-.124-.193-.177-.363-.337-.444-.034-.016-.104-.008-.2.022q-.094.03-.216.087-.06.028-.125.063l-.13.074q-.067.04-.136.086a3 3 0 0 0-.135.096 3 3 0 0 0-.26.219 2 2 0 0 0-.12.121 2 2 0 0 0-.106.128l-.002.002a2 2 0 0 0-.09.132l-.001.001a1.2 1.2 0 0 0-.105.212q-.013.036-.024.073c-1.139-.875-1.563-2.317-1.563-3.203 0-.334.109-.457.305-.454m.836 10.354c.824-1.19.766-2.082-.365-3.194-1.13-1.112-1.789-2.738-1.789-2.738s-.246-.945-.806-.858-.97 1.499.202 2.362c1.173.864-.233 1.45-.685.64-.45-.812-1.683-2.896-2.322-3.295s-1.089-.175-.938.647 2.822 2.813 2.562 3.244-1.176-.506-1.176-.506-2.866-2.567-3.49-1.898.473 1.23 2.037 2.16c1.564.932 1.686 1.178 1.464 1.53s-3.675-2.511-4-1.297c-.323 1.214 3.524 1.567 3.287 2.405-.238.839-2.71-1.587-3.216-.642-.506.946 3.49 2.056 3.522 2.064 1.29.33 4.568 1.028 5.713-.624m5.349 0c-.824-1.19-.766-2.082.365-3.194 1.13-1.112 1.789-2.738 1.789-2.738s.246-.945.806-.858.97 1.499-.202 2.362c-1.173.864.233 1.45.685.64.451-.812 1.683-2.896 2.322-3.295s1.089-.175.938.647-2.822 2.813-2.562 3.244 1.176-.506 1.176-.506 2.866-2.567 3.49-1.898-.473 1.23-2.037 2.16c-1.564.932-1.686 1.178-1.464 1.53s3.675-2.511 4-1.297c.323 1.214-3.524 1.567-3.287 2.405.238.839 2.71-1.587 3.216-.642.506.946-3.49 2.056-3.522 2.064-1.29.33-4.568 1.028-5.713-.624"
     />
   </svg>
+{:else if family === "step"}
+  <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
+    <path
+      d="M22.012 0h1.032v.927H24v.968h-.956V3.78h-1.032V1.896h-1.878v-.97h1.878V0zM2.6 12.371V1.87h.969v10.502h-.97zm10.423.66h10.95v.918h-6.208v9.579h-4.742V13.03zM5.629 3.333v12.356H0v4.51h10.386V8L20.859 8l-.003-4.668-15.227.001z"
+    />
+  </svg>
 {:else}
   <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
     <path

dashboard/src/lib/components/FamilySidebar.svelte

@@ -41,13 +41,13 @@
 </script>
 
 <div
-  class="flex flex-col gap-1 py-2 px-1 border-r border-exo-yellow/10 bg-exo-medium-gray/30 min-w-[64px] overflow-y-auto scrollbar-hide"
+  class="flex flex-col gap-1 py-2 px-1 border-r border-exo-yellow/10 bg-exo-medium-gray/30 min-w-[72px] sm:min-w-[64px] overflow-y-auto scrollbar-hide"
 >
   <!-- All models (no filter) -->
   <button
     type="button"
     onclick={() => onSelect(null)}
-    class="group flex flex-col items-center justify-center p-2 rounded transition-all duration-200 cursor-pointer {selectedFamily ===
+    class="group flex flex-col items-center justify-center p-2 sm:p-2 rounded transition-all duration-200 cursor-pointer min-h-[44px] sm:min-h-0 {selectedFamily ===
     null
       ? 'bg-exo-yellow/20 border-l-2 border-exo-yellow'
       : 'hover:bg-white/5 border-l-2 border-transparent'}"

dashboard/src/lib/components/HeaderNav.svelte

@@ -6,6 +6,12 @@
   export let showSidebarToggle = false;
   export let sidebarVisible = true;
   export let onToggleSidebar: (() => void) | null = null;
+  export let showMobileMenuToggle = false;
+  export let mobileMenuOpen = false;
+  export let onToggleMobileMenu: (() => void) | null = null;
+  export let showMobileRightToggle = false;
+  export let mobileRightOpen = false;
+  export let onToggleMobileRight: (() => void) | null = null;
   export let downloadProgress: {
     count: number;
     percentage: number;
@@ -27,49 +33,96 @@
       onToggleSidebar();
     }
   }
+
+  function handleToggleMobileMenu(): void {
+    if (onToggleMobileMenu) {
+      onToggleMobileMenu();
+    }
+  }
+
+  function handleToggleMobileRight(): void {
+    if (onToggleMobileRight) {
+      onToggleMobileRight();
+    }
+  }
 </script>
 
 <header
-  class="relative z-20 flex items-center justify-center px-6 pt-8 pb-4 bg-exo-dark-gray"
+  class="relative z-20 flex items-center justify-center px-4 md:px-6 pt-4 md:pt-8 pb-3 md:pb-4 bg-exo-dark-gray"
 >
-  <!-- Left: Sidebar Toggle -->
-  {#if showSidebarToggle}
-    <div class="absolute left-6 top-1/2 -translate-y-1/2">
-      <button
-        onclick={handleToggleSidebar}
-        class="p-2 rounded border border-exo-light-gray/30 hover:border-exo-yellow/50 hover:bg-exo-medium-gray/30 transition-colors cursor-pointer"
-        title={sidebarVisible ? "Hide sidebar" : "Show sidebar"}
-        aria-label={sidebarVisible
-          ? "Hide conversation sidebar"
-          : "Show conversation sidebar"}
-        aria-pressed={sidebarVisible}
+  <!-- Left: Sidebar Toggle (desktop) or Mobile Sidebar Toggle (mobile) -->
+  <div
+    class="absolute left-4 md:left-6 top-1/2 -translate-y-1/2 flex items-center gap-2"
+  >
+    <!-- Mobile sidebar toggle -->
+    <button
+      onclick={handleToggleMobileMenu}
+      class="p-2 rounded border border-exo-light-gray/30 hover:border-exo-yellow/50 hover:bg-exo-medium-gray/30 transition-colors cursor-pointer md:hidden"
+      title={mobileMenuOpen ? "Hide sidebar" : "Show sidebar"}
+      aria-label={mobileMenuOpen
+        ? "Hide conversation sidebar"
+        : "Show conversation sidebar"}
+      aria-pressed={mobileMenuOpen}
+    >
+      <svg
+        fill="none"
+        viewBox="0 0 24 24"
+        stroke="currentColor"
+        stroke-width="2"
+        class="w-5 h-5 {mobileMenuOpen
+          ? 'text-exo-yellow'
+          : 'text-exo-light-gray'}"
       >
-        <svg
-          class="w-5 h-5 {sidebarVisible
-            ? 'text-exo-yellow'
-            : 'text-exo-light-gray'}"
-          fill="none"
-          viewBox="0 0 24 24"
-          stroke="currentColor"
-          stroke-width="2"
-        >
-          {#if sidebarVisible}
-            <path
-              stroke-linecap="round"
-              stroke-linejoin="round"
-              d="M11 19l-7-7 7-7m8 14l-7-7 7-7"
-            />
-          {:else}
-            <path
-              stroke-linecap="round"
-              stroke-linejoin="round"
-              d="M13 5l7 7-7 7M5 5l7 7-7 7"
-            />
-          {/if}
-        </svg>
-      </button>
-    </div>
-  {/if}
+        {#if mobileMenuOpen}
+          <path
+            stroke-linecap="round"
+            stroke-linejoin="round"
+            d="M11 19l-7-7 7-7m8 14l-7-7 7-7"
+          ></path>
+        {:else}
+          <path
+            stroke-linecap="round"
+            stroke-linejoin="round"
+            d="M13 5l7 7-7 7M5 5l7 7-7 7"
+          ></path>
+        {/if}
+      </svg>
+    </button>
+    <!-- Desktop sidebar toggle -->
+    <button
+      onclick={handleToggleSidebar}
+      class="p-2 rounded border border-exo-light-gray/30 hover:border-exo-yellow/50 hover:bg-exo-medium-gray/30 transition-colors cursor-pointer hidden md:block"
+      title={sidebarVisible ? "Hide sidebar" : "Show sidebar"}
+      aria-label={sidebarVisible
+        ? "Hide conversation sidebar"
+        : "Show conversation sidebar"}
+      aria-pressed={sidebarVisible}
+    >
+      <svg
+        fill="none"
+        viewBox="0 0 24 24"
+        stroke="currentColor"
+        stroke-width="2"
+        class="w-5 h-5 {sidebarVisible
+          ? 'text-exo-yellow'
+          : 'text-exo-light-gray'}"
+      >
+        {#if sidebarVisible}
+          <path
+            stroke-linecap="round"
+            stroke-linejoin="round"
+            d="M11 19l-7-7 7-7m8 14l-7-7 7-7"
+          ></path>
+        {:else}
+          <path
+            stroke-linecap="round"
+            stroke-linejoin="round"
+            d="M13 5l7 7-7 7M5 5l7 7-7 7"
+          ></path>
+        {/if}
+      </svg>
+    </button>
+  </div>
 
   <!-- Center: Logo (clickable to go home) -->
   <button
@@ -83,19 +136,55 @@
     <img
       src="/exo-logo.png"
       alt="EXO"
-      class="h-18 drop-shadow-[0_0_4px_rgba(255,215,0,0.3)]"
+      class="h-12 md:h-18 drop-shadow-[0_0_4px_rgba(255,215,0,0.3)]"
     />
   </button>
 
-  <!-- Right: Home + Downloads -->
+  <!-- Right: Home + Downloads + Mobile Right Toggle -->
   <nav
-    class="absolute right-6 top-1/2 -translate-y-1/2 flex items-center gap-4"
+    class="absolute right-4 md:right-6 top-1/2 -translate-y-1/2 flex items-center gap-2 md:gap-4"
     aria-label="Main navigation"
   >
+    <!-- Mobile right sidebar toggle (instances/models) - only show when not in chat mode -->
+    {#if showMobileRightToggle}
+      <button
+        onclick={handleToggleMobileRight}
+        class="p-2 rounded border border-exo-light-gray/30 hover:border-exo-yellow/50 hover:bg-exo-medium-gray/30 transition-colors cursor-pointer md:hidden"
+        title={mobileRightOpen ? "Hide instances" : "Show instances"}
+        aria-label={mobileRightOpen
+          ? "Hide instances panel"
+          : "Show instances panel"}
+        aria-pressed={mobileRightOpen}
+      >
+        <svg
+          fill="none"
+          viewBox="0 0 24 24"
+          stroke="currentColor"
+          stroke-width="2"
+          class="w-5 h-5 {mobileRightOpen
+            ? 'text-exo-yellow'
+            : 'text-exo-light-gray'}"
+        >
+          {#if mobileRightOpen}
+            <path
+              stroke-linecap="round"
+              stroke-linejoin="round"
+              d="M13 5l7 7-7 7M5 5l7 7-7 7"
+            ></path>
+          {:else}
+            <path
+              stroke-linecap="round"
+              stroke-linejoin="round"
+              d="M11 19l-7-7 7-7m8 14l-7-7 7-7"
+            ></path>
+          {/if}
+        </svg>
+      </button>
+    {/if}
     {#if showHome}
       <button
         onclick={handleHome}
-        class="text-sm text-white/70 hover:text-exo-yellow transition-colors tracking-wider uppercase flex items-center gap-2 cursor-pointer"
+        class="flex text-sm text-white/70 hover:text-exo-yellow transition-colors tracking-wider uppercase items-center gap-2 cursor-pointer"
         title="Back to topology view"
       >
         <svg
@@ -111,12 +200,12 @@
             d="M3 12l2-2m0 0l7-7 7 7M5 10v10a1 1 0 001 1h3m10-11l2 2m-2-2v10a1 1 0 01-1 1h-3m-6 0a1 1 0 001-1v-4a1 1 0 011-1h2a1 1 0 011 1v4a1 1 0 001 1m-6 0h6"
           />
         </svg>
-        Home
+        <span class="hidden sm:inline">Home</span>
       </button>
     {/if}
     <a
       href="/#/downloads"
-      class="text-sm text-white/70 hover:text-exo-yellow transition-colors tracking-wider uppercase flex items-center gap-2 cursor-pointer"
+      class="text-xs md:text-sm text-white/70 hover:text-exo-yellow transition-colors tracking-wider uppercase flex items-center gap-1.5 md:gap-2 cursor-pointer"
       title="View downloads overview"
     >
       {#if downloadProgress}
@@ -168,7 +257,7 @@
           <path d="M5 21h14" />
         </svg>
       {/if}
-      Downloads
+      <span class="hidden sm:inline">Downloads</span>
     </a>
   </nav>
 </header>

dashboard/src/lib/components/HuggingFaceResultItem.svelte

@@ -36,8 +36,12 @@
     return num.toString();
   }
 
-  // Extract model name from full ID (e.g., "mlx-community/Llama-3.2-1B" -> "Llama-3.2-1B")
-  const modelName = $derived(model.id.split("/").pop() || model.id);
+  // Show short name for mlx-community models, full ID for everything else
+  const modelName = $derived(
+    model.author === "mlx-community"
+      ? model.id.split("/").pop() || model.id
+      : model.id,
+  );
 </script>
 
 <div

dashboard/src/lib/components/MarkdownContent.svelte

@@ -507,9 +507,29 @@
   });
 
   $effect(() => {
-    if (containerRef && processedHtml) {
-      setupCopyButtons();
+    if (!containerRef || !browser) return;
+
+    function handleDelegatedClick(event: MouseEvent) {
+      const codeBtn = (event.target as HTMLElement).closest(
+        ".copy-code-btn",
+      ) as HTMLButtonElement | null;
+      if (codeBtn) {
+        handleCopyClick({ currentTarget: codeBtn } as unknown as Event);
+        return;
+      }
+      const mathBtn = (event.target as HTMLElement).closest(
+        ".copy-math-btn",
+      ) as HTMLButtonElement | null;
+      if (mathBtn) {
+        handleMathCopyClick({ currentTarget: mathBtn } as unknown as Event);
+        return;
+      }
     }
+
+    containerRef.addEventListener("click", handleDelegatedClick);
+    return () => {
+      containerRef?.removeEventListener("click", handleDelegatedClick);
+    };
   });
 </script>
 

dashboard/src/lib/components/ModelPickerGroup.svelte

@@ -32,6 +32,7 @@
     group: ModelGroup;
     isExpanded: boolean;
     isFavorite: boolean;
+    isHighlighted?: boolean;
     selectedModelId: string | null;
     canModelFit: (id: string) => boolean;
     getModelFitStatus: (id: string) => ModelFitStatus;
@@ -48,6 +49,7 @@
     group,
     isExpanded,
     isFavorite,
+    isHighlighted = false,
     selectedModelId,
     canModelFit,
     getModelFitStatus,
@@ -150,10 +152,11 @@
 </script>
 
 <div
+  data-model-ids={group.variants.map((v) => v.id).join(" ")}
   class="border-b border-white/5 last:border-b-0 {!anyVariantFits &&
   !anyVariantHasInstance
     ? 'opacity-50'
-    : ''}"
+    : ''} {isHighlighted ? 'model-just-added' : ''}"
 >
   <!-- Main row -->
   <div
@@ -644,3 +647,21 @@
     </div>
   {/if}
 </div>
+
+<style>
+  .model-just-added {
+    animation: highlightFade 4s ease-out forwards;
+  }
+
+  @keyframes highlightFade {
+    0%,
+    40% {
+      background-color: rgba(20, 83, 45, 0.25);
+      box-shadow: inset 0 0 0 1px rgba(74, 222, 128, 0.4);
+    }
+    100% {
+      background-color: transparent;
+      box-shadow: none;
+    }
+  }
+</style>

dashboard/src/lib/components/ModelPickerModal.svelte

@@ -1,4 +1,5 @@
 <script lang="ts">
+  import { tick } from "svelte";
   import { fade, fly } from "svelte/transition";
   import { cubicOut } from "svelte/easing";
   import FamilySidebar from "./FamilySidebar.svelte";
@@ -7,6 +8,7 @@
   import HuggingFaceResultItem from "./HuggingFaceResultItem.svelte";
   import { getNodesWithModelDownloaded } from "$lib/utils/downloads";
   import { getRecentEntries } from "$lib/stores/recents.svelte";
+  import { addToast } from "$lib/stores/toast.svelte";
 
   interface ModelInfo {
     id: string;
@@ -191,6 +193,13 @@
   let hfSearchDebounceTimer: ReturnType<typeof setTimeout> | null = null;
   let manualModelId = $state("");
   let addModelError = $state<string | null>(null);
+  let justAddedModelId = $state<string | null>(null);
+  let justAddedTimer: ReturnType<typeof setTimeout> | null = null;
+
+  // Inline HuggingFace search in main search bar
+  let mainSearchHfResults = $state<HuggingFaceModel[]>([]);
+  let mainSearchHfLoading = $state(false);
+  let mainSearchDebounceTimer: ReturnType<typeof setTimeout> | null = null;
 
   // Reset transient state when modal opens, but preserve tab selection
   $effect(() => {
@@ -200,6 +209,11 @@
       showFilters = false;
       manualModelId = "";
       addModelError = null;
+      justAddedModelId = null;
+      if (justAddedTimer) {
+        clearTimeout(justAddedTimer);
+        justAddedTimer = null;
+      }
     }
   });
 
@@ -214,6 +228,50 @@
     }
   });
 
+  // Inline HuggingFace search when local search returns no results
+  $effect(() => {
+    const query = searchQuery.trim();
+    const noLocalResults = filteredGroups.length === 0;
+
+    if (mainSearchDebounceTimer) {
+      clearTimeout(mainSearchDebounceTimer);
+      mainSearchDebounceTimer = null;
+    }
+
+    if (
+      selectedFamily === "huggingface" ||
+      selectedFamily === "recents" ||
+      selectedFamily === "favorites" ||
+      query.length < 2 ||
+      !noLocalResults
+    ) {
+      mainSearchHfResults = [];
+      mainSearchHfLoading = false;
+      return;
+    }
+
+    mainSearchHfLoading = true;
+    mainSearchDebounceTimer = setTimeout(async () => {
+      try {
+        const response = await fetch(
+          `/models/search?query=${encodeURIComponent(query)}&limit=10`,
+        );
+        if (response.ok) {
+          const results: HuggingFaceModel[] = await response.json();
+          mainSearchHfResults = results.filter(
+            (r) => !existingModelIds.has(r.id),
+          );
+        } else {
+          mainSearchHfResults = [];
+        }
+      } catch {
+        mainSearchHfResults = [];
+      } finally {
+        mainSearchHfLoading = false;
+      }
+    }, 500);
+  });
+
   async function fetchTrendingModels() {
     hfIsLoadingTrending = true;
     try {
@@ -274,6 +332,24 @@
     addModelError = null;
     try {
       await onAddModel(modelId);
+      // Success: show toast, switch to All Models, highlight the model
+      const shortName = modelId.split("/").pop() || modelId;
+      addToast({ type: "success", message: `Added ${shortName}` });
+      justAddedModelId = modelId;
+      selectedFamily = null;
+      searchQuery = "";
+      // Scroll to the newly added model after DOM update
+      await tick();
+      const el = document.querySelector(
+        `[data-model-ids~="${CSS.escape(modelId)}"]`,
+      );
+      el?.scrollIntoView({ behavior: "smooth", block: "center" });
+      // Clear highlight after 4 seconds
+      if (justAddedTimer) clearTimeout(justAddedTimer);
+      justAddedTimer = setTimeout(() => {
+        justAddedModelId = null;
+        justAddedTimer = null;
+      }, 4000);
     } catch (error) {
       addModelError =
         error instanceof Error ? error.message : "Failed to add model";
@@ -841,6 +917,8 @@
                 {group}
                 isExpanded={expandedGroups.has(group.id)}
                 isFavorite={favorites.has(group.id)}
+                isHighlighted={justAddedModelId !== null &&
+                  group.variants.some((v) => v.id === justAddedModelId)}
                 {selectedModelId}
                 {canModelFit}
                 {getModelFitStatus}
@@ -910,6 +988,8 @@
               {group}
               isExpanded={expandedGroups.has(group.id)}
               isFavorite={favorites.has(group.id)}
+              isHighlighted={justAddedModelId !== null &&
+                group.variants.some((v) => v.id === justAddedModelId)}
               {selectedModelId}
               {canModelFit}
               {getModelFitStatus}
@@ -937,6 +1017,8 @@
               {group}
               isExpanded={expandedGroups.has(group.id)}
               isFavorite={favorites.has(group.id)}
+              isHighlighted={justAddedModelId !== null &&
+                group.variants.some((v) => v.id === justAddedModelId)}
               {selectedModelId}
               {canModelFit}
               {getModelFitStatus}
@@ -948,6 +1030,55 @@
               {instanceStatuses}
             />
           {/each}
+          <!-- Inline HuggingFace search results (shown when no local results match) -->
+          {#if filteredGroups.length === 0 && searchQuery.trim().length >= 2 && selectedFamily !== "huggingface" && selectedFamily !== "recents" && selectedFamily !== "favorites"}
+            {#if mainSearchHfLoading}
+              <div
+                class="flex items-center gap-2 px-3 py-2 border-t border-orange-400/20 bg-orange-950/20"
+              >
+                <span
+                  class="w-4 h-4 border-2 border-orange-400 border-t-transparent rounded-full animate-spin"
+                ></span>
+                <span class="text-xs font-mono text-orange-400/60"
+                  >Searching HuggingFace...</span
+                >
+              </div>
+            {:else if mainSearchHfResults.length > 0}
+              <div
+                class="sticky top-0 z-10 flex items-center gap-2 px-3 py-2 bg-orange-950/30 border-y border-orange-400/20 backdrop-blur-sm"
+              >
+                <span
+                  class="text-xs font-mono text-orange-400 tracking-wider uppercase"
+                  >From HuggingFace</span
+                >
+              </div>
+              {#each mainSearchHfResults as model}
+                <HuggingFaceResultItem
+                  {model}
+                  isAdded={existingModelIds.has(model.id)}
+                  isAdding={addingModelId === model.id}
+                  onAdd={() => handleAddModel(model.id)}
+                  onSelect={() => handleSelectHfModel(model.id)}
+                  downloadedOnNodes={downloadsData
+                    ? getNodesWithModelDownloaded(downloadsData, model.id).map(
+                        getNodeName,
+                      )
+                    : []}
+                />
+              {/each}
+              <button
+                type="button"
+                class="w-full px-3 py-2 text-xs font-mono text-orange-400/60 hover:text-orange-400 hover:bg-orange-500/10 transition-colors text-center"
+                onclick={() => {
+                  hfSearchQuery = searchQuery;
+                  searchHuggingFace(searchQuery);
+                  selectedFamily = "huggingface";
+                }}
+              >
+                See all results on Hub
+              </button>
+            {/if}
+          {/if}
         {/if}
       </div>
     </div>

dashboard/src/lib/stores/app.svelte.ts

@@ -580,6 +580,8 @@ class AppStore {
   debugMode = $state(false);
   topologyOnlyMode = $state(false);
   chatSidebarVisible = $state(true); // Shown by default
+  mobileChatSidebarOpen = $state(false); // Mobile drawer state
+  mobileRightSidebarOpen = $state(false); // Mobile right drawer state
 
   // Image generation params
   imageGenerationParams = $state<ImageGenerationParams>({
@@ -1245,6 +1247,30 @@ class AppStore {
     this.saveChatSidebarVisibleToStorage();
   }
 
+  getMobileChatSidebarOpen(): boolean {
+    return this.mobileChatSidebarOpen;
+  }
+
+  setMobileChatSidebarOpen(open: boolean) {
+    this.mobileChatSidebarOpen = open;
+  }
+
+  toggleMobileChatSidebar() {
+    this.mobileChatSidebarOpen = !this.mobileChatSidebarOpen;
+  }
+
+  getMobileRightSidebarOpen(): boolean {
+    return this.mobileRightSidebarOpen;
+  }
+
+  setMobileRightSidebarOpen(open: boolean) {
+    this.mobileRightSidebarOpen = open;
+  }
+
+  toggleMobileRightSidebar() {
+    this.mobileRightSidebarOpen = !this.mobileRightSidebarOpen;
+  }
+
   startPolling() {
     this.fetchState();
     this.fetchInterval = setInterval(() => this.fetchState(), 1000);
@@ -1551,6 +1577,7 @@ class AppStore {
     // Remove messages after user message (including the user message for image requests
     // since generateImage/editImage will re-add it)
     this.messages = this.messages.slice(0, lastUserIndex);
+    this.updateActiveConversation();
 
     switch (requestType) {
       case "image-generation":
@@ -3158,6 +3185,23 @@ class AppStore {
     return (await response.json()) as TraceStatsResponse;
   }
 
+  /**
+   * Delete traces by task IDs
+   */
+  async deleteTraces(
+    taskIds: string[],
+  ): Promise<{ deleted: string[]; notFound: string[] }> {
+    const response = await fetch("/v1/traces/delete", {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ taskIds }),
+    });
+    if (!response.ok) {
+      throw new Error(`Failed to delete traces: ${response.status}`);
+    }
+    return await response.json();
+  }
+
   /**
    * Get the URL for the raw trace file (for Perfetto)
    */
@@ -3265,6 +3309,18 @@ export const toggleChatSidebarVisible = () =>
   appStore.toggleChatSidebarVisible();
 export const setChatSidebarVisible = (visible: boolean) =>
   appStore.setChatSidebarVisible(visible);
+
+// Mobile sidebar state
+export const mobileChatSidebarOpen = () => appStore.mobileChatSidebarOpen;
+export const toggleMobileChatSidebar = () => appStore.toggleMobileChatSidebar();
+export const setMobileChatSidebarOpen = (open: boolean) =>
+  appStore.setMobileChatSidebarOpen(open);
+export const mobileRightSidebarOpen = () => appStore.mobileRightSidebarOpen;
+export const toggleMobileRightSidebar = () =>
+  appStore.toggleMobileRightSidebar();
+export const setMobileRightSidebarOpen = (open: boolean) =>
+  appStore.setMobileRightSidebarOpen(open);
+
 export const refreshState = () => appStore.fetchState();
 
 // Connection status
@@ -3301,3 +3357,5 @@ export const fetchTraceStats = (taskId: string) =>
   appStore.fetchTraceStats(taskId);
 export const getTraceRawUrl = (taskId: string) =>
   appStore.getTraceRawUrl(taskId);
+export const deleteTraces = (taskIds: string[]) =>
+  appStore.deleteTraces(taskIds);

dashboard/src/routes/+page.svelte

@@ -42,6 +42,9 @@
     setSelectedChatModel,
     selectedChatModel,
     sendMessage,
+    generateImage,
+    editImage,
+    editingImage,
     messages,
     debugMode,
     toggleDebugMode,
@@ -49,6 +52,12 @@
     toggleTopologyOnlyMode,
     chatSidebarVisible,
     toggleChatSidebarVisible,
+    mobileChatSidebarOpen,
+    toggleMobileChatSidebar,
+    setMobileChatSidebarOpen,
+    mobileRightSidebarOpen,
+    toggleMobileRightSidebar,
+    setMobileRightSidebarOpen,
     nodeThunderbolt,
     nodeRdmaCtl,
     thunderboltBridgeCycles,
@@ -79,6 +88,8 @@
   const debugEnabled = $derived(debugMode());
   const topologyOnlyEnabled = $derived(topologyOnlyMode());
   const sidebarVisible = $derived(chatSidebarVisible());
+  const mobileChatOpen = $derived(mobileChatSidebarOpen());
+  const mobileRightOpen = $derived(mobileRightSidebarOpen());
   const tbBridgeCycles = $derived(thunderboltBridgeCycles());
   const tbBridgeData = $derived(nodeThunderboltBridge());
   const identitiesData = $derived(nodeIdentities());
@@ -826,6 +837,52 @@
     if (!model?.tasks) return false;
     return model.tasks.includes("ImageToImage");
   }
+
+  // Route a message to the correct endpoint based on model capabilities.
+  // Image models go to generateImage/editImage; text models go to sendMessage.
+  function routeMessage(
+    content: string,
+    files?: {
+      id: string;
+      name: string;
+      type: string;
+      textContent?: string;
+      preview?: string;
+    }[],
+  ) {
+    const model = selectedChatModel();
+    if (!model) {
+      sendMessage(content, files, null);
+      return;
+    }
+
+    const currentEditImage = editingImage();
+
+    // Image editing mode (explicit edit or attached image with ImageToImage model)
+    if (currentEditImage && content && modelSupportsImageEditing(model)) {
+      editImage(content, currentEditImage.imageDataUrl);
+      return;
+    }
+    if (
+      modelSupportsImageEditing(model) &&
+      files?.length &&
+      files[0].preview &&
+      content
+    ) {
+      editImage(content, files[0].preview);
+      return;
+    }
+
+    // Text-to-image generation
+    if (modelSupportsImageGeneration(model) && content) {
+      generateImage(content);
+      return;
+    }
+
+    // Default: text chat
+    sendMessage(content, files, null);
+  }
+
   let selectedSharding = $state<"Pipeline" | "Tensor">("Pipeline");
   type InstanceMeta = "MlxRing" | "MlxJaccl";
 
@@ -1519,7 +1576,11 @@
           downloadKind
         ] as Record<string, unknown>;
 
-        if (downloadKind !== "DownloadOngoing") continue;
+        if (
+          downloadKind !== "DownloadOngoing" &&
+          downloadKind !== "DownloadPending"
+        )
+          continue;
         if (!downloadPayload) continue;
 
         const downloadModelId = extractModelIdFromDownload(downloadPayload);
@@ -1534,9 +1595,38 @@
           if (downloadModelId !== modelId) continue;
         }
 
-        isDownloading = true;
+        // For DownloadPending with partial bytes (paused/resumed downloads),
+        // synthesize a progress object from the top-level downloaded/total fields
+        let progress: DownloadProgress | null;
+        if (downloadKind === "DownloadPending") {
+          const pendingDownloaded = getBytes(
+            downloadPayload.downloaded ??
+              downloadPayload.downloaded_bytes ??
+              downloadPayload.downloadedBytes,
+          );
+          const pendingTotal = getBytes(
+            downloadPayload.total ??
+              downloadPayload.total_bytes ??
+              downloadPayload.totalBytes,
+          );
+          if (pendingDownloaded <= 0 && pendingTotal <= 0) continue;
+          isDownloading = true;
+          progress = {
+            totalBytes: pendingTotal,
+            downloadedBytes: pendingDownloaded,
+            speed: 0,
+            etaMs: 0,
+            percentage:
+              pendingTotal > 0 ? (pendingDownloaded / pendingTotal) * 100 : 0,
+            completedFiles: 0,
+            totalFiles: 0,
+            files: [],
+          };
+        } else {
+          isDownloading = true;
+          progress = parseDownloadProgress(downloadPayload);
+        }
 
-        const progress = parseDownloadProgress(downloadPayload);
         if (progress) {
           // Sum all values across nodes - each node downloads independently
           totalBytes += progress.totalBytes;
@@ -1688,7 +1778,11 @@
           }
         }
 
-        if (downloadKind !== "DownloadOngoing") continue;
+        if (
+          downloadKind !== "DownloadOngoing" &&
+          downloadKind !== "DownloadPending"
+        )
+          continue;
         if (!downloadPayload) continue;
 
         // Check if this download is for this instance's model
@@ -1698,9 +1792,37 @@
           downloadModelId &&
           downloadModelId === instanceModelId
         ) {
-          isDownloading = true;
+          // For DownloadPending with partial bytes, synthesize progress
+          let progress: DownloadProgress | null;
+          if (downloadKind === "DownloadPending") {
+            const pendingDownloaded = getBytes(
+              downloadPayload.downloaded ??
+                downloadPayload.downloaded_bytes ??
+                downloadPayload.downloadedBytes,
+            );
+            const pendingTotal = getBytes(
+              downloadPayload.total ??
+                downloadPayload.total_bytes ??
+                downloadPayload.totalBytes,
+            );
+            if (pendingDownloaded <= 0 && pendingTotal <= 0) continue;
+            isDownloading = true;
+            progress = {
+              totalBytes: pendingTotal,
+              downloadedBytes: pendingDownloaded,
+              speed: 0,
+              etaMs: 0,
+              percentage:
+                pendingTotal > 0 ? (pendingDownloaded / pendingTotal) * 100 : 0,
+              completedFiles: 0,
+              totalFiles: 0,
+              files: [],
+            };
+          } else {
+            isDownloading = true;
+            progress = parseDownloadProgress(downloadPayload);
+          }
 
-          const progress = parseDownloadProgress(downloadPayload);
           if (progress) {
             // Sum all values across nodes - each node downloads independently
             totalBytes += progress.totalBytes;
@@ -2778,7 +2900,7 @@
         // Running model is same or better tier — use it directly
         setSelectedChatModel(bestRunning.id);
         if (!chatStarted) createConversation();
-        sendMessage(content, files);
+        routeMessage(content, files);
         return;
       }
     }
@@ -2795,7 +2917,7 @@
     if (hasRunningInstance(autoModel.id)) {
       setSelectedChatModel(autoModel.id);
       if (!chatStarted) createConversation();
-      sendMessage(content, files);
+      routeMessage(content, files);
       return;
     }
 
@@ -2948,7 +3070,7 @@
       if (pendingAutoMessage) {
         const msg = pendingAutoMessage;
         pendingAutoMessage = null;
-        sendMessage(msg.content, msg.files);
+        routeMessage(msg.content, msg.files);
       }
       return;
     }
@@ -3027,7 +3149,7 @@
     // Model is selected and running — send directly
     if (model && hasRunningInstance(model)) {
       chatLaunchState = "ready";
-      sendMessage(content, files, null);
+      routeMessage(content, files);
       return;
     }
 
@@ -4604,16 +4726,35 @@
       showSidebarToggle={true}
       {sidebarVisible}
       onToggleSidebar={toggleChatSidebarVisible}
+      showMobileMenuToggle={true}
+      mobileMenuOpen={mobileChatOpen}
+      onToggleMobileMenu={toggleMobileChatSidebar}
+      showMobileRightToggle={!chatStarted && !topologyOnlyEnabled}
+      {mobileRightOpen}
+      onToggleMobileRight={toggleMobileRightSidebar}
       downloadProgress={activeDownloadSummary}
     />
   {/if}
 
+  <!-- Mobile Chat Sidebar Drawer -->
+  {#if !topologyOnlyEnabled}
+    <ChatSidebar
+      isMobileDrawer={true}
+      isOpen={mobileChatOpen}
+      onClose={() => setMobileChatSidebarOpen(false)}
+      onNewChat={handleNewChat}
+      onSelectConversation={() => {
+        userForcedIdle = false;
+      }}
+    />
+  {/if}
+
   <!-- Main Content -->
   <main class="flex-1 flex overflow-hidden relative">
-    <!-- Left: Conversation History Sidebar (hidden in topology-only mode, welcome state, or when toggled off) -->
+    <!-- Left: Conversation History Sidebar (hidden in topology-only mode, welcome state, or when toggled off) - Desktop only -->
     {#if !topologyOnlyEnabled && sidebarVisible}
       <div
-        class="w-80 flex-shrink-0 border-r border-exo-yellow/10"
+        class="hidden md:block w-80 flex-shrink-0 border-r border-exo-yellow/10"
         role="complementary"
         aria-label="Conversation history"
       >
@@ -4923,11 +5064,33 @@
           </div>
         </div>
 
-        <!-- Right Sidebar: Instance Controls (wider on welcome page for better visibility) -->
+        <!-- Mobile Right Sidebar Drawer (Instances) -->
+        {#if mobileRightOpen}
+          <!-- Overlay backdrop -->
+          <button
+            type="button"
+            class="fixed inset-0 bg-black/60 backdrop-blur-sm z-40 md:hidden"
+            onclick={() => setMobileRightSidebarOpen(false)}
+            aria-label="Close instances panel"
+          ></button>
+          <!-- Drawer panel -->
+          <aside
+            class="fixed right-0 top-0 bottom-0 w-80 bg-exo-dark-gray border-l border-exo-yellow/10 z-50 flex flex-col md:hidden overflow-y-auto"
+            aria-label="Instance controls mobile"
+          >
+            {@render rightSidebarContent()}
+          </aside>
+        {/if}
+
+        <!-- Right Sidebar: Instance Controls (wider on welcome page for better visibility) - Desktop only -->
         <aside
-          class="w-80 border-l border-exo-yellow/10 bg-exo-dark-gray flex flex-col flex-shrink-0"
+          class="hidden md:flex w-80 border-l border-exo-yellow/10 bg-exo-dark-gray flex-col flex-shrink-0"
           aria-label="Instance controls"
         >
+          {@render rightSidebarContent()}
+        </aside>
+
+        {#snippet rightSidebarContent()}
           <!-- Running Instances Panel (only shown when instances exist) - Scrollable -->
           {#if instanceCount > 0}
             <div class="p-4 flex-shrink-0">
@@ -5809,7 +5972,7 @@
               {/if}
             </div>
           </div>
-        </aside>
+        {/snippet}
       </div>
     {:else}
       <!-- CHAT STATE: Chat + Mini-Map -->
@@ -6022,10 +6185,10 @@
           {/if}
         </div>
 
-        <!-- Right: Mini-Map Sidebar -->
+        <!-- Right: Mini-Map Sidebar - Desktop only -->
         {#if minimized}
           <aside
-            class="w-80 border-l border-exo-yellow/20 bg-exo-dark-gray flex flex-col flex-shrink-0 overflow-y-auto"
+            class="hidden md:flex w-80 border-l border-exo-yellow/20 bg-exo-dark-gray flex-col flex-shrink-0 overflow-y-auto"
             in:fly={{ x: 100, duration: 400, easing: cubicInOut }}
             aria-label="Cluster topology"
           >
@@ -6050,12 +6213,11 @@
               </div>
 
               <div
-                class="relative aspect-square bg-exo-dark-gray rounded-lg overflow-hidden"
+                class="relative aspect-square bg-exo-dark-gray rounded-lg overflow-hidden pointer-events-none"
               >
                 <TopologyGraph
                   highlightedNodes={highlightedNodes()}
                   filteredNodes={nodeFilter}
-                  onNodeClick={togglePreviewNodeFilter}
                 />
 
                 {@render clusterWarningsCompact()}

dashboard/src/routes/traces/+page.svelte

@@ -3,6 +3,7 @@
   import {
     listTraces,
     getTraceRawUrl,
+    deleteTraces,
     type TraceListItem,
   } from "$lib/stores/app.svelte";
   import HeaderNav from "$lib/components/HeaderNav.svelte";
@@ -10,6 +11,51 @@
   let traces = $state<TraceListItem[]>([]);
   let loading = $state(true);
   let error = $state<string | null>(null);
+  let selectedIds = $state<Set<string>>(new Set());
+  let deleting = $state(false);
+
+  let allSelected = $derived(
+    traces.length > 0 && selectedIds.size === traces.length,
+  );
+
+  function toggleSelect(taskId: string) {
+    const next = new Set(selectedIds);
+    if (next.has(taskId)) {
+      next.delete(taskId);
+    } else {
+      next.add(taskId);
+    }
+    selectedIds = next;
+  }
+
+  function toggleSelectAll() {
+    if (allSelected) {
+      selectedIds = new Set();
+    } else {
+      selectedIds = new Set(traces.map((t) => t.taskId));
+    }
+  }
+
+  async function handleDelete() {
+    if (selectedIds.size === 0) return;
+    const count = selectedIds.size;
+    if (
+      !confirm(
+        `Delete ${count} trace${count === 1 ? "" : "s"}? This cannot be undone.`,
+      )
+    )
+      return;
+    deleting = true;
+    try {
+      await deleteTraces([...selectedIds]);
+      selectedIds = new Set();
+      await refresh();
+    } catch (e) {
+      error = e instanceof Error ? e.message : "Failed to delete traces";
+    } finally {
+      deleting = false;
+    }
+  }
 
   function formatBytes(bytes: number): string {
     if (!bytes || bytes <= 0) return "0B";
@@ -109,6 +155,16 @@
         </h1>
       </div>
       <div class="flex items-center gap-3">
+        {#if selectedIds.size > 0}
+          <button
+            type="button"
+            class="text-xs font-mono text-red-400 hover:text-red-300 transition-colors uppercase border border-red-500/40 px-2 py-1 rounded"
+            onclick={handleDelete}
+            disabled={deleting}
+          >
+            {deleting ? "Deleting..." : `Delete (${selectedIds.size})`}
+          </button>
+        {/if}
         <button
           type="button"
           class="text-xs font-mono text-exo-light-gray hover:text-exo-yellow transition-colors uppercase border border-exo-medium-gray/40 px-2 py-1 rounded"
@@ -143,14 +199,41 @@
       </div>
     {:else}
       <div class="space-y-3">
+        <div class="flex items-center gap-2 px-1">
+          <button
+            type="button"
+            class="text-xs font-mono uppercase transition-colors {allSelected
+              ? 'text-exo-yellow'
+              : 'text-exo-light-gray hover:text-exo-yellow'}"
+            onclick={toggleSelectAll}
+          >
+            {allSelected ? "Deselect all" : "Select all"}
+          </button>
+        </div>
         {#each traces as trace}
+          {@const isSelected = selectedIds.has(trace.taskId)}
+          <!-- svelte-ignore a11y_no_static_element_interactions -->
           <div
-            class="rounded border border-exo-medium-gray/30 bg-exo-black/30 p-4 flex items-center justify-between gap-4"
+            role="button"
+            tabindex="0"
+            class="w-full text-left rounded border-l-2 border-r border-t border-b transition-all p-4 flex items-center justify-between gap-4 cursor-pointer {isSelected
+              ? 'bg-exo-yellow/10 border-l-exo-yellow border-r-exo-medium-gray/30 border-t-exo-medium-gray/30 border-b-exo-medium-gray/30'
+              : 'bg-exo-black/30 border-l-transparent border-r-exo-medium-gray/30 border-t-exo-medium-gray/30 border-b-exo-medium-gray/30 hover:bg-white/[0.03]'}"
+            onclick={() => toggleSelect(trace.taskId)}
+            onkeydown={(e) => {
+              if (e.key === "Enter" || e.key === " ") {
+                e.preventDefault();
+                toggleSelect(trace.taskId);
+              }
+            }}
           >
             <div class="min-w-0 flex-1">
               <a
                 href="#/traces/{trace.taskId}"
-                class="text-sm font-mono text-white hover:text-exo-yellow transition-colors truncate block"
+                class="text-sm font-mono transition-colors truncate block {isSelected
+                  ? 'text-exo-yellow'
+                  : 'text-white hover:text-exo-yellow'}"
+                onclick={(e) => e.stopPropagation()}
               >
                 {trace.taskId}
               </a>
@@ -160,7 +243,11 @@
                 )}
               </div>
             </div>
-            <div class="flex items-center gap-2 shrink-0">
+            <!-- svelte-ignore a11y_click_events_have_key_events -->
+            <div
+              class="flex items-center gap-2 shrink-0"
+              onclick={(e) => e.stopPropagation()}
+            >
               <a
                 href="#/traces/{trace.taskId}"
                 class="text-xs font-mono text-exo-light-gray hover:text-exo-yellow transition-colors uppercase border border-exo-medium-gray/40 px-2 py-1 rounded"

docs/api.md

@@ -4,7 +4,7 @@ This document describes the REST API exposed by the **EXO** service, as implemen
 
 `src/exo/master/api.py`
 
-The API is used to manage model instances in the cluster, inspect cluster state, and perform inference using an OpenAI-compatible interface.
+The API is used to manage model instances in the cluster, inspect cluster state, and perform inference using multiple API-compatible interfaces.
 
 Base URL example:
 
@@ -144,8 +144,59 @@ Placement result.
 
 Returns the list of available models and their metadata.
 
+**Query parameters:**
+
+* `status`: string (optional) - Filter by `downloaded` to show only downloaded models
+
 **Response:**
-Array of model descriptors.
+Array of model descriptors including `is_custom` field for custom HuggingFace models.
+
+### Add Custom Model
+
+**POST** `/models/add`
+
+Add a custom model from HuggingFace hub.
+
+**Request body (example):**
+
+```json
+{
+  "model_id": "mlx-community/my-custom-model"
+}
+```
+
+**Response:**
+Model descriptor for the added model.
+
+**Security note:**
+Models with `trust_remote_code` enabled in their configuration require explicit opt-in (default is false) for security.
+
+### Delete Custom Model
+
+**DELETE** `/models/custom/{model_id}`
+
+Delete a user-added custom model card.
+
+**Path parameters:**
+
+* `model_id`: string, ID of the custom model to delete
+
+**Response:**
+Confirmation JSON with deleted model ID.
+
+### Search Models
+
+**GET** `/models/search`
+
+Search HuggingFace Hub for mlx-community models.
+
+**Query parameters:**
+
+* `query`: string (optional) - Search query
+* `limit`: integer (default: 20) - Maximum number of results
+
+**Response:**
+Array of HuggingFace model search results.
 
 ## 4. Inference / Chat Completions
 
@@ -168,9 +219,123 @@ Executes a chat completion request using an OpenAI-compatible schema. Supports s
 }
 ```
 
+**Request parameters:**
+
+* `model`: string, required - Model ID to use
+* `messages`: array, required - Conversation messages
+* `stream`: boolean (default: false) - Enable streaming responses
+* `max_tokens`: integer (optional) - Maximum tokens to generate
+* `temperature`: float (optional) - Sampling temperature
+* `top_p`: float (optional) - Nucleus sampling parameter
+* `top_k`: integer (optional) - Top-k sampling parameter
+* `stop`: string or array (optional) - Stop sequences
+* `seed`: integer (optional) - Random seed for reproducibility
+* `enable_thinking`: boolean (optional) - Enable thinking mode for capable models (DeepSeek V3.1, Qwen3, GLM-4.7)
+* `tools`: array (optional) - Tool definitions for function calling
+* `logprobs`: boolean (optional) - Return log probabilities
+* `top_logprobs`: integer (optional) - Number of top log probabilities to return
+
 **Response:**
 OpenAI-compatible chat completion response.
 
+**Streaming response format:**
+When `stream=true`, returns Server-Sent Events (SSE) with format:
+
+```
+data: {"id":"...","object":"chat.completion","created":...,"model":"...","choices":[...]}
+
+data: [DONE]
+```
+
+**Non-streaming response includes usage statistics:**
+
+```json
+{
+  "id": "...",
+  "object": "chat.completion",
+  "created": 1234567890,
+  "model": "llama-3.2-1b",
+  "choices": [{
+    "index": 0,
+    "message": {
+      "role": "assistant",
+      "content": "Hello! How can I help you?"
+    },
+    "finish_reason": "stop"
+  }],
+  "usage": {
+    "prompt_tokens": 15,
+    "completion_tokens": 8,
+    "total_tokens": 23
+  }
+}
+```
+
+**Cancellation:**
+You can cancel an active generation by closing the HTTP connection. The server detects the disconnection and stops processing.
+
+### Claude Messages API
+
+**POST** `/v1/messages`
+
+Executes a chat completion request using the Claude Messages API format. Supports streaming and non-streaming modes.
+
+**Request body (example):**
+
+```json
+{
+  "model": "llama-3.2-1b",
+  "messages": [
+    { "role": "user", "content": "Hello" }
+  ],
+  "max_tokens": 1024,
+  "stream": false
+}
+```
+
+**Streaming response format:**
+When `stream=true`, returns Server-Sent Events with Claude-specific event types:
+
+* `message_start` - Message generation started
+* `content_block_start` - Content block started
+* `content_block_delta` - Incremental content chunk
+* `content_block_stop` - Content block completed
+* `message_delta` - Message metadata updates
+* `message_stop` - Message generation completed
+
+**Response:**
+Claude-compatible messages response.
+
+### OpenAI Responses API
+
+**POST** `/v1/responses`
+
+Executes a chat completion request using the OpenAI Responses API format. Supports streaming and non-streaming modes.
+
+**Request body (example):**
+
+```json
+{
+  "model": "llama-3.2-1b",
+  "messages": [
+    { "role": "user", "content": "Hello" }
+  ],
+  "stream": false
+}
+```
+
+**Streaming response format:**
+When `stream=true`, returns Server-Sent Events with response-specific event types:
+
+* `response.created` - Response generation started
+* `response.in_progress` - Response is being generated
+* `response.output_item.added` - New output item added
+* `response.output_item.done` - Output item completed
+* `response.done` - Response generation completed
+
+**Response:**
+OpenAI Responses API-compatible response.
+
 ### Benchmarked Chat Completions
 
 **POST** `/bench/chat/completions`
@@ -181,26 +346,177 @@ Same as `/v1/chat/completions`, but also returns performance and generation stat
 Same schema as `/v1/chat/completions`.
 
 **Response:**
-Chat completion plus benchmarking metrics.
+Chat completion plus benchmarking metrics including:
 
-## 5. Image Generation & Editing
+* `prompt_tps` - Tokens per second during prompt processing
+* `generation_tps` - Tokens per second during generation
+* `prompt_tokens` - Number of prompt tokens
+* `generation_tokens` - Number of generated tokens
+* `peak_memory_usage` - Peak memory used during generation
+
+### Cancel Command
+
+**POST** `/v1/cancel/{command_id}`
+
+Cancels an active generation command (text or image). Notifies workers and closes the stream.
+
+**Path parameters:**
+
+* `command_id`: string, ID of the command to cancel
+
+**Response (example):**
+
+```json
+{
+  "message": "Command cancelled.",
+  "command_id": "cmd-abc-123"
+}
+```
+
+Returns 404 if the command is not found or already completed.
+
+## 5. Ollama API Compatibility
+
+EXO provides Ollama API compatibility for tools like OpenWebUI.
+
+### Ollama Chat
+
+**POST** `/ollama/api/chat`
+**POST** `/ollama/api/api/chat` (alias)
+**POST** `/ollama/api/v1/chat` (alias)
+
+Execute a chat request using Ollama API format.
+
+**Request body (example):**
+
+```json
+{
+  "model": "llama-3.2-1b",
+  "messages": [
+    { "role": "user", "content": "Hello" }
+  ],
+  "stream": false
+}
+```
+
+**Response:**
+Ollama-compatible chat response.
+
+### Ollama Generate
+
+**POST** `/ollama/api/generate`
+
+Execute a text generation request using Ollama API format.
+
+**Request body (example):**
+
+```json
+{
+  "model": "llama-3.2-1b",
+  "prompt": "Hello",
+  "stream": false
+}
+```
+
+**Response:**
+Ollama-compatible generation response.
+
+### Ollama Tags
+
+**GET** `/ollama/api/tags`
+**GET** `/ollama/api/api/tags` (alias)
+**GET** `/ollama/api/v1/tags` (alias)
+
+Returns list of downloaded models in Ollama tags format.
+
+**Response:**
+Array of model tags with metadata.
+
+### Ollama Show
+
+**POST** `/ollama/api/show`
+
+Returns model information in Ollama show format.
+
+**Request body:**
+
+```json
+{
+  "name": "llama-3.2-1b"
+}
+```
+
+**Response:**
+Model details including modelfile and family.
+
+### Ollama PS
+
+**GET** `/ollama/api/ps`
+
+Returns list of running models (active instances).
+
+**Response:**
+Array of active model instances.
+
+### Ollama Version
+
+**GET** `/ollama/api/version`
+**HEAD** `/ollama/` (alias)
+**HEAD** `/ollama/api/version` (alias)
+
+Returns version information for Ollama API compatibility.
+
+**Response:**
+
+```json
+{
+  "version": "exo v1.0"
+}
+```
+
+## 6. Image Generation & Editing
 
 ### Image Generation
 
 **POST** `/v1/images/generations`
 
-Executes an image generation request using an OpenAI-compatible schema with additional advanced_params.
+Executes an image generation request using an OpenAI-compatible schema with additional advanced_params. Supports both streaming and non-streaming modes.
 
 **Request body (example):**
 
 ```json
 {
   "prompt": "a robot playing chess",
-  "model": "flux-dev",
+  "model": "exolabs/FLUX.1-dev",
+  "n": 1,
+  "size": "1024x1024",
   "stream": false,
+  "response_format": "b64_json"
 }
 ```
 
+**Request parameters:**
+
+* `prompt`: string, required - Text description of the image
+* `model`: string, required - Image model ID
+* `n`: integer (default: 1) - Number of images to generate
+* `size`: string (default: "auto") - Image dimensions. Supported sizes:
+  - `512x512`
+  - `768x768`
+  - `1024x768`
+  - `768x1024`
+  - `1024x1024`
+  - `1024x1536`
+  - `1536x1024`
+  - `1024x1365`
+  - `1365x1024`
+* `stream`: boolean (default: false) - Enable streaming for partial images
+* `partial_images`: integer (default: 0) - Number of partial images to stream during generation
+* `response_format`: string (default: "b64_json") - Either `url` or `b64_json`
+* `quality`: string (default: "medium") - Either `high`, `medium`, or `low`
+* `output_format`: string (default: "png") - Either `png`, `jpeg`, or `webp`
+* `advanced_params`: object (optional) - Advanced generation parameters
+
 **Advanced Parameters (`advanced_params`):**
 
 | Parameter | Type | Constraints | Description |
@@ -210,8 +526,54 @@ Executes an image generation request using an OpenAI-compatible schema with addi
 | `guidance` | float | 1.0-20.0 | Classifier-free guidance scale |
 | `negative_prompt` | string | - | Text describing what to avoid in the image |
 
+**Non-streaming response:**
+
+```json
+{
+  "created": 1234567890,
+  "data": [
+    {
+      "b64_json": "iVBORw0KGgoAAAANSUhEUgAA...",
+      "url": null
+    }
+  ]
+}
+```
+
+**Streaming response format:**
+When `stream=true` and `partial_images > 0`, returns Server-Sent Events:
+
+```
+data: {"type":"partial","image_index":0,"partial_index":1,"total_partials":5,"format":"png","data":{"b64_json":"..."}}
+
+data: {"type":"final","image_index":0,"format":"png","data":{"b64_json":"..."}}
+
+data: [DONE]
+```
+
+### Image Editing
+
+**POST** `/v1/images/edits`
+
+Executes an image editing request (img2img) using FLUX.1-Kontext-dev or similar models.
+
+**Request (multipart/form-data):**
+
+* `image`: file, required - Input image to edit
+* `prompt`: string, required - Text description of desired changes
+* `model`: string, required - Image editing model ID (e.g., `exolabs/FLUX.1-Kontext-dev`)
+* `n`: integer (default: 1) - Number of edited images to generate
+* `size`: string (optional) - Output image dimensions
+* `response_format`: string (default: "b64_json") - Either `url` or `b64_json`
+* `input_fidelity`: string (default: "low") - Either `low` or `high` - Controls how closely the output follows the input image
+* `stream`: string (default: "false") - Enable streaming
+* `partial_images`: string (default: "0") - Number of partial images to stream
+* `quality`: string (default: "medium") - Either `high`, `medium`, or `low`
+* `output_format`: string (default: "png") - Either `png`, `jpeg`, or `webp`
+* `advanced_params`: string (optional) - JSON-encoded advanced parameters
+
 **Response:**
-OpenAI-compatible image generation response.
+Same format as `/v1/images/generations`.
 
 ### Benchmarked Image Generation
 
@@ -223,16 +585,15 @@ Same as `/v1/images/generations`, but also returns generation statistics.
 Same schema as `/v1/images/generations`.
 
 **Response:**
-Image generation plus benchmarking metrics.
+Image generation plus benchmarking metrics including:
 
-### Image Editing
-
-**POST** `/v1/images/edits`
-
-Executes an image editing request using an OpenAI-compatible schema with additional advanced_params (same as `/v1/images/generations`).
-
-**Response:**
-Same format as `/v1/images/generations`.
+* `seconds_per_step` - Average time per denoising step
+* `total_generation_time` - Total generation time
+* `num_inference_steps` - Number of inference steps used
+* `num_images` - Number of images generated
+* `image_width` - Output image width
+* `image_height` - Output image height
+* `peak_memory_usage` - Peak memory used during generation
 
 ### Benchmarked Image Editing
 
@@ -246,36 +607,127 @@ Same schema as `/v1/images/edits`.
 **Response:**
 Same format as `/bench/images/generations`, including `generation_stats`.
 
-## 6. Complete Endpoint Summary
+### List Images
+
+**GET** `/images`
+
+List all stored images.
+
+**Response:**
+Array of image metadata including URLs and expiration times.
+
+### Get Image
+
+**GET** `/images/{image_id}`
+
+Retrieve a stored image by ID.
+
+**Path parameters:**
+
+* `image_id`: string, ID of the image
+
+**Response:**
+Image file with appropriate content type.
+
+## 7. Complete Endpoint Summary
 
 ```
+# General
 GET     /node_id
 GET     /state
 GET     /events
 
+# Instance Management
 POST    /instance
 GET     /instance/{instance_id}
 DELETE  /instance/{instance_id}
-
 GET     /instance/previews
 GET     /instance/placement
 POST    /place_instance
 
+# Models
 GET     /models
 GET     /v1/models
+POST    /models/add
+DELETE  /models/custom/{model_id}
+GET     /models/search
 
+# Text Generation (OpenAI Chat Completions)
 POST    /v1/chat/completions
 POST    /bench/chat/completions
 
+# Text Generation (Claude Messages API)
+POST    /v1/messages
+
+# Text Generation (OpenAI Responses API)
+POST    /v1/responses
+
+# Text Generation (Ollama API)
+POST    /ollama/api/chat
+POST    /ollama/api/api/chat
+POST    /ollama/api/v1/chat
+POST    /ollama/api/generate
+GET     /ollama/api/tags
+GET     /ollama/api/api/tags
+GET     /ollama/api/v1/tags
+POST    /ollama/api/show
+GET     /ollama/api/ps
+GET     /ollama/api/version
+HEAD    /ollama/
+HEAD    /ollama/api/version
+
+# Command Control
+POST    /v1/cancel/{command_id}
+
+# Image Generation
 POST    /v1/images/generations
 POST    /bench/images/generations
 POST    /v1/images/edits
 POST    /bench/images/edits
+GET     /images
+GET     /images/{image_id}
 ```
 
-## 7. Notes
+## 8. Notes
 
-* The `/v1/chat/completions` endpoint is compatible with the OpenAI Chat API format, so existing OpenAI clients can be pointed to EXO by changing the base URL.
-* The `/v1/images/generations` and `/v1/images/edits` endpoints are compatible with the OpenAI Images API format.
-* The instance placement endpoints allow you to plan and preview cluster allocations before actually creating instances.
-* The `/events` and `/state` endpoints are primarily intended for operational visibility and debugging.
+### API Compatibility
+
+EXO provides multiple API-compatible interfaces:
+
+* **OpenAI Chat Completions API** - Compatible with OpenAI clients and tools
+* **Claude Messages API** - Compatible with Anthropic's Claude API format
+* **OpenAI Responses API** - Compatible with OpenAI's Responses API format
+* **Ollama API** - Compatible with Ollama and tools like OpenWebUI
+
+Existing OpenAI, Claude, or Ollama clients can be pointed to EXO by changing the base URL.
+
+### Custom Models
+
+You can add custom models from HuggingFace using the `/models/add` endpoint. Custom models are identified by the `is_custom` field in model list responses.
+
+**Security:** Models requiring `trust_remote_code` must be explicitly enabled (default is false) for security. Only enable this if you trust the model's remote code.
+
+### Usage Statistics
+
+Chat completion responses include usage statistics with:
+
+* `prompt_tokens` - Number of tokens in the prompt
+* `completion_tokens` - Number of tokens generated
+* `total_tokens` - Sum of prompt and completion tokens
+
+### Request Cancellation
+
+You can cancel active requests by:
+
+1. Closing the HTTP connection (for streaming requests)
+2. Calling `/v1/cancel/{command_id}` (for any request)
+
+The server detects cancellation and stops processing immediately.
+
+### Instance Placement
+
+The instance placement endpoints allow you to plan and preview cluster allocations before creating instances. This helps optimize resource usage across nodes.
+
+### Observability
+
+The `/events` and `/state` endpoints are primarily intended for operational visibility and debugging.

docs/architecture.md

@@ -28,6 +28,26 @@ There are currently 5 major systems:
     
     Implements a distributed algorithm for master election in unstable networking conditions
 
+## API Layer
+
+The API system uses multiple adapters to support multiple API formats, converting them to a single request / response type.
+
+### Adapter Pattern
+
+Adapters convert between external API formats and EXO's internal types:
+
+```
+Chat Completions → [adapter] → TextGenerationTaskParams → Application
+Claude Messages  → [adapter] → TextGenerationTaskParams → Application
+Responses API    → [adapter] → TextGenerationTaskParams → Application
+Ollama API       → [adapter] → TextGenerationTaskParams → Application
+```
+
+Each adapter implements two key functions:
+1. **Request conversion**: Converts API-specific requests to `TextGenerationTaskParams`
+2. **Response generation**: Converts internal `TokenChunk` streams back to API-specific formats (streaming and non-streaming)
+
+
 ## Topics
 
 There are currently 5 topics:

flake.nix

@@ -108,6 +108,7 @@
                 package = pkgsSwift.swiftPackages.swift-format;
               };
               shfmt.enable = true;
+              taplo.enable = true;
             };
           };
 
@@ -116,12 +117,13 @@
               uvLock = builtins.fromTOML (builtins.readFile ./uv.lock);
               mlxPackage = builtins.head (builtins.filter (p: p.name == "mlx" && p.source ? git) uvLock.package);
               uvLockMlxVersion = mlxPackage.version;
+              uvLockMlxRev = builtins.elemAt (builtins.split "#" mlxPackage.source.git) 2;
             in
             {
               metal-toolchain = pkgs.callPackage ./nix/metal-toolchain.nix { };
               mlx = pkgs.callPackage ./nix/mlx.nix {
                 inherit (self'.packages) metal-toolchain;
-                inherit uvLockMlxVersion;
+                inherit uvLockMlxVersion uvLockMlxRev;
               };
               default = self'.packages.exo;
             }

nix/mlx.nix

@@ -11,6 +11,7 @@
 , fmt
 , python313Packages
 , uvLockMlxVersion
+, uvLockMlxRev
 }:
 
 assert stdenv.isDarwin;
@@ -41,15 +42,13 @@ let
 
   mlx = stdenv.mkDerivation rec {
     pname = "mlx";
-    version = let v = "0.30.7.dev20260225+257d5692"; in
-      assert v == uvLockMlxVersion || throw "MLX version mismatch: nix/mlx.nix has ${v} but uv.lock has ${uvLockMlxVersion}. Update both the version and hash in nix/mlx.nix.";
-      v;
+    version = uvLockMlxVersion;
     pyproject = true;
 
     src = fetchFromGitHub {
       owner = "rltakashige";
       repo = "mlx-jaccl-fix-small-recv";
-      rev = "257d5692fc7af6bba3b8afaeb63c549b7d1e43d5";
+      rev = uvLockMlxRev;
       hash = "sha256-GosFIWxIB48Egb1MqJrR3xhsUsQeWdRk5rV93USY6wQ=";
     };
 

pyproject.toml

@@ -5,31 +5,30 @@ description = "Exo"
 readme = "README.md"
 requires-python = ">=3.13"
 dependencies = [
-    "aiofiles>=24.1.0",
-    "aiohttp>=3.12.14",
-    "types-aiofiles>=24.1.0.20250708",
-    "pydantic>=2.11.7",
-    "fastapi>=0.116.1",
-    "filelock>=3.18.0",
-    "rustworkx>=0.17.1",
-    "huggingface-hub>=0.33.4",
-    "psutil>=7.0.0",
-    "loguru>=0.7.3",
-    "exo_pyo3_bindings", # rust bindings
-    "anyio==4.11.0",
-    "mlx; sys_platform == 'darwin'",
-    "mlx[cpu]==0.30.6; sys_platform == 'linux'",
-    "mlx-lm==0.30.7",
-    "tiktoken>=0.12.0", # required for kimi k2 tokenizer
-    "hypercorn>=0.18.0",
-    "openai-harmony>=0.0.8",
-    "httpx>=0.28.1",
-    "tomlkit>=0.14.0",
-    "pillow>=11.0,<12.0", # compatibility with mflux
-    "mflux==0.15.5",
-    "python-multipart>=0.0.21",
-    "msgspec>=0.19.0",
-    "zstandard>=0.23.0",
+  "aiofiles>=24.1.0",
+  "aiohttp>=3.12.14",
+  "types-aiofiles>=24.1.0.20250708",
+  "pydantic>=2.11.7",
+  "fastapi>=0.116.1",
+  "filelock>=3.18.0",
+  "rustworkx>=0.17.1",
+  "huggingface-hub>=0.33.4",
+  "psutil>=7.0.0",
+  "loguru>=0.7.3",
+  "exo_pyo3_bindings",                         # rust bindings
+  "anyio==4.11.0",
+  "mlx; sys_platform == 'darwin'",
+  "mlx[cpu]==0.30.6; sys_platform == 'linux'",
+  "mlx-lm",
+  "tiktoken>=0.12.0",                          # required for kimi k2 tokenizer
+  "hypercorn>=0.18.0",
+  "openai-harmony>=0.0.8",
+  "httpx>=0.28.1",
+  "tomlkit>=0.14.0",
+  "mflux==0.16.9",
+  "python-multipart>=0.0.21",
+  "msgspec>=0.19.0",
+  "zstandard>=0.23.0",
 ]
 
 [project.scripts]
@@ -38,12 +37,12 @@ exo = "exo.main:main"
 # dependencies only required for development
 [dependency-groups]
 dev = [
-    "basedpyright>=1.29.0",
-    "pyinstaller>=6.17.0",
-    "pytest>=8.4.0",
-    "pytest-asyncio>=1.0.0",
-    "pytest-env",
-    "ruff>=0.11.13",
+  "basedpyright>=1.29.0",
+  "pyinstaller>=6.17.0",
+  "pytest>=8.4.0",
+  "pytest-asyncio>=1.0.0",
+  "pytest-env",
+  "ruff>=0.11.13",
 ]
 
 # mlx[cuda] requires a newer version of mlx. the ideal on linux is: default to mlx[cpu] unless[cuda] specified.
@@ -57,15 +56,12 @@ dev = [
 ###
 
 [tool.uv.workspace]
-members = [
-    "rust/exo_pyo3_bindings",
-    "bench",
-]
+members = ["rust/exo_pyo3_bindings", "bench"]
 
 [tool.uv.sources]
 exo_pyo3_bindings = { workspace = true }
 mlx = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git", branch = "address-rdma-gpu-locks", marker = "sys_platform == 'darwin'" }
-#mlx-lm = { git = "https://github.com/davidmcc73/mlx-lm", branch = "stable" }
+mlx-lm = { git = "https://github.com/rltakashige/mlx-lm", branch = "leo/eval-left-padding-in-batched-rotation" }
 # Uncomment to use local mlx/mlx-lm development versions:
 # mlx = { path = "/Users/Shared/mlx", editable=true }
 # mlx-lm = { path = "/Users/Shared/mlx-lm", editable=true }
@@ -95,7 +91,15 @@ reportUnnecessaryTypeIgnoreComment = "error"
 pythonVersion = "3.13"
 pythonPlatform = "Darwin"
 
-exclude = ["**/.venv", "**/venv", "**/__pycache__", "**/exo_scripts", "**/.direnv", "**/rust", "**/.github"]
+exclude = [
+  "**/.venv",
+  "**/venv",
+  "**/__pycache__",
+  "**/exo_scripts",
+  "**/.direnv",
+  "**/rust",
+  "**/.github",
+]
 stubPath = ".mlx_typings"
 
 [[tool.basedpyright.executionEnvironments]]
@@ -109,17 +113,19 @@ root = "src"
 [tool.uv]
 required-version = ">=0.8.6"
 prerelease = "allow"
-environments = [
-    "sys_platform == 'darwin'",
-    "sys_platform == 'linux'",
-]
+environments = ["sys_platform == 'darwin'", "sys_platform == 'linux'"]
 
 ###
 # ruff configuration
 ###
 
 [tool.ruff]
-extend-exclude = ["shared/protobufs/**", "*mlx_typings/**", "rust/exo_pyo3_bindings/**"]
+extend-exclude = [
+  "shared/protobufs/**",
+  "*mlx_typings/**",
+  "rust/exo_pyo3_bindings/**",
+  "bench/vendor/**",
+]
 
 [tool.ruff.lint]
 extend-select = ["I", "N", "B", "A", "PIE", "SIM"]
@@ -127,13 +133,7 @@ extend-select = ["I", "N", "B", "A", "PIE", "SIM"]
 [tool.pytest.ini_options]
 pythonpath = "."
 asyncio_mode = "auto"
-markers = [
-    "slow: marks tests as slow (deselected by default)"
-]
-env = [
-  "EXO_TESTS=1"
-]
+markers = ["slow: marks tests as slow (deselected by default)"]
+env = ["EXO_TESTS=1"]
 addopts = "-m 'not slow' --ignore=tests/start_distributed_test.py"
-filterwarnings = [
-    "ignore:builtin type Swig:DeprecationWarning",
-]
+filterwarnings = ["ignore:builtin type Swig:DeprecationWarning"]

python/parts.nix

@@ -43,6 +43,27 @@
             final.setuptools
           ];
         });
+        # rouge-score and sacrebleu don't declare setuptools as a build dependency
+        rouge-score = prev.rouge-score.overrideAttrs (old: {
+          nativeBuildInputs = (old.nativeBuildInputs or [ ]) ++ [
+            final.setuptools
+          ];
+        });
+        sacrebleu = prev.sacrebleu.overrideAttrs (old: {
+          nativeBuildInputs = (old.nativeBuildInputs or [ ]) ++ [
+            final.setuptools
+          ];
+        });
+        sqlitedict = prev.sqlitedict.overrideAttrs (old: {
+          nativeBuildInputs = (old.nativeBuildInputs or [ ]) ++ [
+            final.setuptools
+          ];
+        });
+        word2number = prev.word2number.overrideAttrs (old: {
+          nativeBuildInputs = (old.nativeBuildInputs or [ ]) ++ [
+            final.setuptools
+          ];
+        });
       } // lib.optionalAttrs pkgs.stdenv.hostPlatform.isDarwin {
         # Use our pure Nix-built MLX with Metal support (macOS only)
         mlx = self'.packages.mlx;
@@ -96,13 +117,16 @@
         "lib/python3.13/site-packages/nvidia*"
       ];
 
-      exoVenv = (pythonSet.mkVirtualEnv "exo-env" workspace.deps.default).overrideAttrs {
+      # Exclude bench deps from main env (bench has its own benchVenv)
+      exoDeps = removeAttrs workspace.deps.default [ "exo-bench" ];
+
+      exoVenv = (pythonSet.mkVirtualEnv "exo-env" exoDeps).overrideAttrs {
         venvIgnoreCollisions = venvCollisionPaths;
       };
 
       # Virtual environment with dev dependencies for testing
       testVenv = (pythonSet.mkVirtualEnv "exo-test-env" (
-        workspace.deps.default // {
+        exoDeps // {
           exo = [ "dev" ]; # Include pytest, pytest-asyncio, pytest-env
         }
       )).overrideAttrs {
@@ -158,6 +182,7 @@
           exo-test-env = testVenv;
         } // {
         exo-bench = mkBenchScript "exo-bench" (inputs.self + /bench/exo_bench.py);
+        exo-eval = mkBenchScript "exo-eval" (inputs.self + /bench/exo_eval.py);
         exo-eval-tool-calls = mkBenchScript "exo-eval-tool-calls" (inputs.self + /bench/eval_tool_calls.py);
         exo-get-all-models-on-cluster = mkSimplePythonScript "exo-get-all-models-on-cluster" (inputs.self + /tests/get_all_models_on_cluster.py);
       };

resources/inference_model_cards/mlx-community--DeepSeek-V3.1-4bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/DeepSeek-V3.1-4bit"
 n_layers = 61
 hidden_size = 7168
+num_key_value_heads = 128
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "deepseek"

resources/inference_model_cards/mlx-community--DeepSeek-V3.1-8bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/DeepSeek-V3.1-8bit"
 n_layers = 61
 hidden_size = 7168
+num_key_value_heads = 128
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "deepseek"

resources/inference_model_cards/mlx-community--GLM-4.5-Air-8bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/GLM-4.5-Air-8bit"
 n_layers = 46
 hidden_size = 4096
+num_key_value_heads = 8
 supports_tensor = false
 tasks = ["TextGeneration"]
 family = "glm"

resources/inference_model_cards/mlx-community--GLM-4.5-Air-bf16.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/GLM-4.5-Air-bf16"
 n_layers = 46
 hidden_size = 4096
+num_key_value_heads = 8
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "glm"

resources/inference_model_cards/mlx-community--GLM-4.7-4bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/GLM-4.7-4bit"
 n_layers = 91
 hidden_size = 5120
+num_key_value_heads = 8
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "glm"

resources/inference_model_cards/mlx-community--GLM-4.7-6bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/GLM-4.7-6bit"
 n_layers = 91
 hidden_size = 5120
+num_key_value_heads = 8
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "glm"

resources/inference_model_cards/mlx-community--GLM-4.7-8bit-gs32.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/GLM-4.7-8bit-gs32"
 n_layers = 91
 hidden_size = 5120
+num_key_value_heads = 8
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "glm"

resources/inference_model_cards/mlx-community--GLM-4.7-Flash-4bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/GLM-4.7-Flash-4bit"
 n_layers = 47
 hidden_size = 2048
+num_key_value_heads = 20
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "glm"

resources/inference_model_cards/mlx-community--GLM-4.7-Flash-5bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/GLM-4.7-Flash-5bit"
 n_layers = 47
 hidden_size = 2048
+num_key_value_heads = 20
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "glm"

resources/inference_model_cards/mlx-community--GLM-4.7-Flash-6bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/GLM-4.7-Flash-6bit"
 n_layers = 47
 hidden_size = 2048
+num_key_value_heads = 20
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "glm"

resources/inference_model_cards/mlx-community--GLM-4.7-Flash-8bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/GLM-4.7-Flash-8bit"
 n_layers = 47
 hidden_size = 2048
+num_key_value_heads = 20
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "glm"

resources/inference_model_cards/mlx-community--GLM-5-8bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/GLM-5-8bit-MXFP8"
 n_layers = 78
 hidden_size = 6144
+num_key_value_heads = 64
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "glm"

resources/inference_model_cards/mlx-community--GLM-5-MXFP4-Q8.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/GLM-5-MXFP4-Q8"
 n_layers = 78
 hidden_size = 6144
+num_key_value_heads = 64
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "glm"

resources/inference_model_cards/mlx-community--GLM-5-bf16.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/GLM-5"
 n_layers = 78
 hidden_size = 6144
+num_key_value_heads = 64
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "glm"

resources/inference_model_cards/mlx-community--Kimi-K2-Instruct-4bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Kimi-K2-Instruct-4bit"
 n_layers = 61
 hidden_size = 7168
+num_key_value_heads = 64
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "kimi"

resources/inference_model_cards/mlx-community--Kimi-K2-Thinking.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Kimi-K2-Thinking"
 n_layers = 61
 hidden_size = 7168
+num_key_value_heads = 64
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "kimi"

resources/inference_model_cards/mlx-community--Kimi-K2.5.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Kimi-K2.5"
 n_layers = 61
 hidden_size = 7168
+num_key_value_heads = 64
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "kimi"

resources/inference_model_cards/mlx-community--Llama-3.1-Nemotron-70B-Instruct-HF-4bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Llama-3.1-Nemotron-70B-Instruct-HF-4bit"
+n_layers = 80
+hidden_size = 8192
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "llama"
+quantization = "4bit"
+base_model = "NVIDIA Llama-3.1-Nemotron-70B-Instruct"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 39688355840

resources/inference_model_cards/mlx-community--Llama-3.1-Nemotron-70B-Instruct-HF-8bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Llama-3.1-Nemotron-70B-Instruct-HF-8bit"
+n_layers = 80
+hidden_size = 8192
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "llama"
+quantization = "8bit"
+base_model = "NVIDIA Llama-3.1-Nemotron-70B-Instruct"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 74964549632

resources/inference_model_cards/mlx-community--Llama-3.1-Nemotron-70B-Instruct-HF-bf16.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Llama-3.1-Nemotron-70B-Instruct-HF-bf16"
+n_layers = 80
+hidden_size = 8192
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "llama"
+quantization = "bf16"
+base_model = "NVIDIA Llama-3.1-Nemotron-70B-Instruct"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 141107412992

resources/inference_model_cards/mlx-community--Llama-3.1-Nemotron-Nano-4B-v1.1-4bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Llama-3.1-Nemotron-Nano-4B-v1.1-4bit"
+n_layers = 32
+hidden_size = 3072
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "llama"
+quantization = "4bit"
+base_model = "NVIDIA Llama-3.1-Nemotron-Nano-4B-v1.1"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 2538706944

resources/inference_model_cards/mlx-community--Llama-3.1-Nemotron-Nano-4B-v1.1-8bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Llama-3.1-Nemotron-Nano-4B-v1.1-8bit"
+n_layers = 32
+hidden_size = 3072
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "llama"
+quantization = "8bit"
+base_model = "NVIDIA Llama-3.1-Nemotron-Nano-4B-v1.1"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 4794980352

resources/inference_model_cards/mlx-community--Llama-3.1-Nemotron-Nano-4B-v1.1-bf16.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Llama-3.1-Nemotron-Nano-4B-v1.1-bf16"
+n_layers = 32
+hidden_size = 3072
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "llama"
+quantization = "bf16"
+base_model = "NVIDIA Llama-3.1-Nemotron-Nano-4B-v1.1"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 9025492992

resources/inference_model_cards/mlx-community--Llama-3.2-1B-Instruct-4bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Llama-3.2-1B-Instruct-4bit"
 n_layers = 16
 hidden_size = 2048
+num_key_value_heads = 8
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "llama"

resources/inference_model_cards/mlx-community--Llama-3.2-3B-Instruct-4bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Llama-3.2-3B-Instruct-4bit"
 n_layers = 28
 hidden_size = 3072
+num_key_value_heads = 8
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "llama"

resources/inference_model_cards/mlx-community--Llama-3.2-3B-Instruct-8bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Llama-3.2-3B-Instruct-8bit"
 n_layers = 28
 hidden_size = 3072
+num_key_value_heads = 8
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "llama"

resources/inference_model_cards/mlx-community--Llama-3.3-70B-Instruct-4bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Llama-3.3-70B-Instruct-4bit"
 n_layers = 80
 hidden_size = 8192
+num_key_value_heads = 8
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "llama"

resources/inference_model_cards/mlx-community--Llama-3.3-70B-Instruct-8bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Llama-3.3-70B-Instruct-8bit"
 n_layers = 80
 hidden_size = 8192
+num_key_value_heads = 8
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "llama"

resources/inference_model_cards/mlx-community--Meta-Llama-3.1-70B-Instruct-4bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Meta-Llama-3.1-70B-Instruct-4bit"
 n_layers = 80
 hidden_size = 8192
+num_key_value_heads = 8
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "llama"

resources/inference_model_cards/mlx-community--Meta-Llama-3.1-8B-Instruct-4bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Meta-Llama-3.1-8B-Instruct-4bit"
 n_layers = 32
 hidden_size = 4096
+num_key_value_heads = 8
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "llama"

resources/inference_model_cards/mlx-community--Meta-Llama-3.1-8B-Instruct-8bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Meta-Llama-3.1-8B-Instruct-8bit"
 n_layers = 32
 hidden_size = 4096
+num_key_value_heads = 8
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "llama"

resources/inference_model_cards/mlx-community--Meta-Llama-3.1-8B-Instruct-bf16.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Meta-Llama-3.1-8B-Instruct-bf16"
 n_layers = 32
 hidden_size = 4096
+num_key_value_heads = 8
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "llama"

resources/inference_model_cards/mlx-community--MiniMax-M2.1-3bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/MiniMax-M2.1-3bit"
 n_layers = 61
 hidden_size = 3072
+num_key_value_heads = 8
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "minimax"

resources/inference_model_cards/mlx-community--MiniMax-M2.1-8bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/MiniMax-M2.1-8bit"
 n_layers = 61
 hidden_size = 3072
+num_key_value_heads = 8
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "minimax"

resources/inference_model_cards/mlx-community--MiniMax-M2.5-4bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/MiniMax-M2.5-4bit"
 n_layers = 62
 hidden_size = 3072
+num_key_value_heads = 8
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "minimax"

resources/inference_model_cards/mlx-community--MiniMax-M2.5-6bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/MiniMax-M2.5-6bit"
 n_layers = 62
 hidden_size = 3072
+num_key_value_heads = 8
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "minimax"

resources/inference_model_cards/mlx-community--MiniMax-M2.5-8bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/MiniMax-M2.5-8bit"
 n_layers = 62
 hidden_size = 3072
+num_key_value_heads = 8
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "minimax"

resources/inference_model_cards/mlx-community--NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-4Bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-4Bit"
+n_layers = 52
+hidden_size = 2688
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "nemotron"
+quantization = "4bit"
+base_model = "NVIDIA Nemotron-3-Nano-30B-A3B"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 17775342336

resources/inference_model_cards/mlx-community--NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-5Bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-5Bit"
+n_layers = 52
+hidden_size = 2688
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "nemotron"
+quantization = "5bit"
+base_model = "NVIDIA Nemotron-3-Nano-30B-A3B"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 21721476864

resources/inference_model_cards/mlx-community--NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-6Bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-6Bit"
+n_layers = 52
+hidden_size = 2688
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "nemotron"
+quantization = "6bit"
+base_model = "NVIDIA Nemotron-3-Nano-30B-A3B"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 25667611392

resources/inference_model_cards/mlx-community--NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-8Bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-8Bit"
+n_layers = 52
+hidden_size = 2688
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "nemotron"
+quantization = "8bit"
+base_model = "NVIDIA Nemotron-3-Nano-30B-A3B"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 33559880448

resources/inference_model_cards/mlx-community--NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-BF16.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-BF16"
+n_layers = 52
+hidden_size = 2688
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "nemotron"
+quantization = "bf16"
+base_model = "NVIDIA Nemotron-3-Nano-30B-A3B"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 63155889408

resources/inference_model_cards/mlx-community--NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-MXFP4.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/NVIDIA-Nemotron-3-Nano-30B-A3B-MLX-MXFP4"
+n_layers = 52
+hidden_size = 2688
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "nemotron"
+quantization = "4bit"
+base_model = "NVIDIA Nemotron-3-Nano-30B-A3B"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 16788808704

resources/inference_model_cards/mlx-community--NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4"
+n_layers = 52
+hidden_size = 2688
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "nemotron"
+quantization = "4bit"
+base_model = "NVIDIA Nemotron-3-Nano-30B-A3B"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 19323906944

resources/inference_model_cards/mlx-community--NVIDIA-Nemotron-Nano-9B-v2-4bits.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/NVIDIA-Nemotron-Nano-9B-v2-4bits"
+n_layers = 56
+hidden_size = 4480
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "nemotron"
+quantization = "4bit"
+base_model = "NVIDIA Nemotron-Nano-9B-v2"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 5002791936

resources/inference_model_cards/mlx-community--NVIDIA-Nemotron-Nano-9B-v2-6bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/NVIDIA-Nemotron-Nano-9B-v2-6bit"
+n_layers = 56
+hidden_size = 4480
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "nemotron"
+quantization = "6bit"
+base_model = "NVIDIA Nemotron-Nano-9B-v2"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 7224298496

resources/inference_model_cards/mlx-community--Qwen3-0.6B-4bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Qwen3-0.6B-4bit"
 n_layers = 28
 hidden_size = 1024
+num_key_value_heads = 8
 supports_tensor = false
 tasks = ["TextGeneration"]
 family = "qwen"

resources/inference_model_cards/mlx-community--Qwen3-0.6B-8bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Qwen3-0.6B-8bit"
 n_layers = 28
 hidden_size = 1024
+num_key_value_heads = 8
 supports_tensor = false
 tasks = ["TextGeneration"]
 family = "qwen"

resources/inference_model_cards/mlx-community--Qwen3-235B-A22B-Instruct-2507-4bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Qwen3-235B-A22B-Instruct-2507-4bit"
 n_layers = 94
 hidden_size = 4096
+num_key_value_heads = 4
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "qwen"

resources/inference_model_cards/mlx-community--Qwen3-235B-A22B-Instruct-2507-8bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Qwen3-235B-A22B-Instruct-2507-8bit"
 n_layers = 94
 hidden_size = 4096
+num_key_value_heads = 4
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "qwen"

resources/inference_model_cards/mlx-community--Qwen3-30B-A3B-4bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Qwen3-30B-A3B-4bit"
 n_layers = 48
 hidden_size = 2048
+num_key_value_heads = 4
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "qwen"

resources/inference_model_cards/mlx-community--Qwen3-30B-A3B-8bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Qwen3-30B-A3B-8bit"
 n_layers = 48
 hidden_size = 2048
+num_key_value_heads = 4
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "qwen"

resources/inference_model_cards/mlx-community--Qwen3-Coder-480B-A35B-Instruct-4bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Qwen3-Coder-480B-A35B-Instruct-4bit"
 n_layers = 62
 hidden_size = 6144
+num_key_value_heads = 8
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "qwen"

resources/inference_model_cards/mlx-community--Qwen3-Coder-480B-A35B-Instruct-8bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Qwen3-Coder-480B-A35B-Instruct-8bit"
 n_layers = 62
 hidden_size = 6144
+num_key_value_heads = 8
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "qwen"

resources/inference_model_cards/mlx-community--Qwen3-Coder-Next-4bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Qwen3-Coder-Next-4bit"
 n_layers = 48
 hidden_size = 2048
+num_key_value_heads = 2
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "qwen"

resources/inference_model_cards/mlx-community--Qwen3-Coder-Next-5bit.toml

@@ -1,6 +1,7 @@
 model_id = "mlx-community/Qwen3-Coder-Next-5bit"
 n_layers = 48
 hidden_size = 2048
+num_key_value_heads = 2
 supports_tensor = true
 tasks = ["TextGeneration"]
 family = "qwen"