try using different fork

pyproject.toml

@@ -66,7 +66,7 @@ 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/rltakashige/mlx-lm", branch = "leo/eval-left-padding-in-batched-rotation" }
 torch = [{ index = "pytorch-cu130", marker = "sys_platform == 'linux'" }]
-vllm = { git = "https://github.com/hmellor/vllm.git", branch = "transformers-v5" }
+vllm = { git = "https://github.com/rltakashige/vllm-transformers-sm121.git" }
 # 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 }

src/exo/worker/engines/vllm/growable_cache.py

@@ -26,7 +26,6 @@ def patch_vllm() -> None:
     _patch_allocate_slots()
     _patch_get_computed_blocks()
     _patch_moe_sum()
-    _patch_marlin_w2_thread_config()
     logger.info("vLLM growable KV cache patch applied")
 
 
@@ -275,22 +274,6 @@ def _patch_moe_sum() -> None:
     ops.moe_sum = moe_sum_f32  # type: ignore
 
 
-def _patch_marlin_w2_thread_config() -> None:
-    try:
-        import vllm._custom_ops as ops  # type: ignore[reportMissingImports]
-    except ImportError:
-        return
-
-    original_gemm = ops.moe_wna16_marlin_gemm
-
-    def patched_gemm(*args: "object", **kwargs: "object") -> "object":
-        kwargs["thread_k"] = 64
-        kwargs["thread_n"] = 128
-        return original_gemm(*args, **kwargs)
-
-    ops.moe_wna16_marlin_gemm = patched_gemm  # type: ignore
-
-
 def _patch_get_computed_blocks() -> None:
     from vllm.v1.core.kv_cache_manager import KVCacheBlocks, KVCacheManager
     from vllm.v1.core.kv_cache_utils import KVCacheBlock