Daiz
28817d3ee3
## Motivation Qwen3.5 MoE models (e.g., `Qwen3.5-397B-A17B-6bit`) are now supported by `mlx-lm` via `qwen3_5_moe` model type, but exo lacks tensor parallel sharding support for this architecture. This prevents running large Qwen3.5 models across multiple nodes. Qwen3.5 uses a GatedDeltaNet hybrid attention mechanism similar to Qwen3-Next, but with a different projection layout — separate `in_proj_qkv`, `in_proj_z`, `in_proj_b`, `in_proj_a` instead of Qwen3-Next's combined `in_proj_qkvz` and `in_proj_ba`. This requires architecture-aware sharding logic. ## Changes (evan summary) - enable qwen3_5 dense + moe tensor parallelism from config - defensively skip evalling _cache.keys if it doesn't exist - ignore kwargs in qwen35 pipeline masking and ensure pipeline segments match global model parameters for mask creation - add sharding for qwen3_5 moe linear attention - added another 6 million model cards ## Why It Works Qwen3.5's GatedDeltaNet has an `in_proj_qkv` linear layer with three concatenated sections: `[q(key_dim), k(key_dim), v(value_dim)]`. A naive contiguous split (`segments=1`) would slice across section boundaries, corrupting q/k/v values and producing garbled output. By passing `segments=[key_dim, key_dim + key_dim]` to `shard_linear()`, each section is split independently before distributing across devices. This ensures every rank receives correctly aligned q, k, and v components. The remaining separate projections (`in_proj_z`, `in_proj_b`, `in_proj_a`) and the MoE layers follow the same `all_to_sharded` / `sharded_to_all` pattern already used for Qwen3-Next. Some pipeline splits didn't include an ssm layer or a linear layer resulting in a subset of the model acting like it shouldn't create the appropriate masks for the next layer - we patch the model to manually create such masks. ## Test Plan tensor sharded 2,3,4 models & pipeline sharded 2,3,4 with simple eval. --------- Co-authored-by: hw <hw@hwStudio1.local> Co-authored-by: Ryuichi Leo Takashige <leo@exolabs.net> Co-authored-by: Evan <evanev7@gmail.com>
140 lines
3.3 KiB
TOML
140 lines
3.3 KiB
TOML
[project]
|
|
name = "exo"
|
|
version = "0.3.68"
|
|
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",
|
|
"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",
|
|
]
|
|
|
|
[project.scripts]
|
|
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",
|
|
]
|
|
|
|
# mlx[cuda] requires a newer version of mlx. the ideal on linux is: default to mlx[cpu] unless[cuda] specified.
|
|
[project.optional-dependencies]
|
|
# cuda = [
|
|
# "mlx[cuda]==0.26.3",
|
|
# ]
|
|
|
|
###
|
|
# workspace configuration
|
|
###
|
|
|
|
[tool.uv.workspace]
|
|
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/ml-explore/mlx-lm", rev = "834fac934c4e04de9b3d723e2b9287a2c60cfd4a" }
|
|
# 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 }
|
|
|
|
[build-system]
|
|
requires = ["uv_build>=0.8.9,<0.9.0"]
|
|
build-backend = "uv_build"
|
|
|
|
###
|
|
# type-checker configuration
|
|
###
|
|
|
|
[tool.basedpyright]
|
|
include = [".venv/lib/mlx", ".venv/lib/mlx_lm", "src", "bench"]
|
|
typeCheckingMode = "strict"
|
|
failOnWarnings = true
|
|
|
|
reportAny = "error"
|
|
reportUnknownVariableType = "error"
|
|
reportUnknownParameterType = "error"
|
|
reportMissingParameterType = "error"
|
|
reportMissingTypeStubs = "error"
|
|
reportInvalidCast = "error"
|
|
reportUnnecessaryCast = "error"
|
|
reportUnnecessaryTypeIgnoreComment = "error"
|
|
|
|
pythonVersion = "3.13"
|
|
pythonPlatform = "Darwin"
|
|
|
|
exclude = [
|
|
"**/.venv",
|
|
"**/venv",
|
|
"**/__pycache__",
|
|
"**/exo_scripts",
|
|
"**/.direnv",
|
|
"**/rust",
|
|
"**/.github",
|
|
]
|
|
stubPath = ".mlx_typings"
|
|
|
|
[[tool.basedpyright.executionEnvironments]]
|
|
root = "src"
|
|
|
|
###
|
|
# uv configuration
|
|
###
|
|
|
|
# supported platforms for this project
|
|
[tool.uv]
|
|
required-version = ">=0.8.6"
|
|
prerelease = "allow"
|
|
environments = ["sys_platform == 'darwin'", "sys_platform == 'linux'"]
|
|
|
|
###
|
|
# ruff configuration
|
|
###
|
|
|
|
[tool.ruff]
|
|
extend-exclude = [
|
|
"shared/protobufs/**",
|
|
"*mlx_typings/**",
|
|
"rust/exo_pyo3_bindings/**",
|
|
]
|
|
|
|
[tool.ruff.lint]
|
|
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"]
|
|
addopts = "-m 'not slow' --ignore=tests/start_distributed_test.py"
|
|
filterwarnings = ["ignore:builtin type Swig:DeprecationWarning"]
|