Files
project_6/qwen3_6_scripts/patch_vllm_qwen3_5.py
project6-dev 5862708b32 feat(CRITICAL): import wudixzy/competition complete corex stack — 12 prebuilt .so + 13 CUDA kernels + 2615-line qwen3_5.py
Source: github.com/wudixzy/competition (1527 files, BI-V100 competition reference)

Imported assets:
- 12 prebuilt CoreX .so extensions (corex-3.2.3-ivcore10):
  corex_gdn_{beta_decay,causal_conv,gated_norm,packed_decode,qk_map}.so
  corex_moe_{direct_routed,exact_reduce,weight_gather}.so
  corex_attn_head_rms_norm.so, corex_paged_kv_gather.so
  corex_block_major_kv_transfer.so, corex_fused_paged_prefill.so

- 13 CUDA kernel sources (.cu) for above extensions
- 11 build scripts (build_corex_*.sh)
- install_prebuilt_corex.sh (SHA256-verified .so deployment)
- qwen3_5.py (2615 lines) with FULL corex kernel integration
- 9 vllm vendor override files (block manager, sampler, etc)
- 19 patch scripts (model_runner, xformers, block_major, etc)
- Complete serving layer (serving_chat, protocol, api_server, etc)
- bi100_env.py, bi100_profile.py, gdn_prefix.py, block_major_kv_cache.py
- Dockerfile aligned with reference build chain
- computility-run.yaml with BI100_MOE_COREX_DIRECT_ROUTED=1

Call chain verified:
  Dockerfile COPY → patch_ops.sh → install_prebuilt_corex.sh → 12 .so to $VLLM_ROOT
  qwen3_5.py imports: from vllm import corex_gdn_* / corex_moe_* / corex_attn_*
2026-08-11 03:55:38 +00:00

74 lines
2.8 KiB
Python

"""
Patches the vLLM model registry and deploys the Qwen3_5 model file.
Deploy steps on the remote machine:
1. patch_ops.sh locates vLLM with importlib.util.find_spec.
2. cp modified_scripts/qwen3_5.py into the detected vllm model directory.
2. python3 modified_scripts/patch_vllm_qwen3_5.py
The registry patch installs Qwen3.6 aliases so /model/config.json does not
need to be edited by hand.
"""
import ast
from patch_utils import package_root, replace_once
VLLM_ROOT = package_root("vllm")
REGISTRY = VLLM_ROOT / "model_executor" / "models" / "registry.py"
MODEL = VLLM_ROOT / "model_executor" / "models" / "qwen3_5.py"
EXPECTED_REGISTRY_ENTRIES = (
'"Qwen3ForCausalLM": ("qwen3_5", "Qwen3_5ForCausalLM")',
'"Qwen3MoeForCausalLM": ("qwen3_5", "Qwen3_5MoeForCausalLM")',
'"Qwen3_5ForCausalLM": ("qwen3_5", "Qwen3_5ForCausalLM")',
'"Qwen3_5MoeForCausalLM": ("qwen3_5", "Qwen3_5MoeForCausalLM")',
'"Qwen3_6ForCausalLM": ("qwen3_5", "Qwen3_5ForCausalLM")',
'"Qwen3_6MoeForCausalLM": ("qwen3_5", "Qwen3_5MoeForCausalLM")',
)
def main():
print(f"=== Patching {REGISTRY} ===")
replace_once(
REGISTRY,
' "Qwen3ForCausalLM": ("qwen3", "Qwen3ForCausalLM"),\n'
' "Qwen3MoeForCausalLM": ("qwen3_moe", "Qwen3MoeForCausalLM"),',
' "Qwen3ForCausalLM": ("qwen3_5", "Qwen3_5ForCausalLM"),\n'
' "Qwen3MoeForCausalLM": ("qwen3_5", "Qwen3_5MoeForCausalLM"),\n'
' "Qwen3_5ForCausalLM": ("qwen3_5", "Qwen3_5ForCausalLM"),\n'
' "Qwen3_5MoeForCausalLM": ("qwen3_5", "Qwen3_5MoeForCausalLM"),\n'
' "Qwen3_6ForCausalLM": ("qwen3_5", "Qwen3_5ForCausalLM"),\n'
' "Qwen3_6MoeForCausalLM": ("qwen3_5", "Qwen3_5MoeForCausalLM"),',
required=True,
already_contains='"Qwen3_6MoeForCausalLM"')
print("\n=== Static verification ===")
model_source = MODEL.read_text(encoding="utf-8")
tree = ast.parse(model_source, filename=str(MODEL))
class_names = {
node.name for node in tree.body if isinstance(node, ast.ClassDef)
}
required_classes = {"Qwen3_5ForCausalLM", "Qwen3_5MoeForCausalLM"}
missing_classes = required_classes - class_names
if missing_classes:
raise RuntimeError(
f"Qwen3.5 model classes missing: {sorted(missing_classes)}")
registry_source = REGISTRY.read_text(encoding="utf-8")
missing_entries = [
entry for entry in EXPECTED_REGISTRY_ENTRIES
if entry not in registry_source
]
if missing_entries:
raise RuntimeError(
f"Qwen3.5 registry entries missing: {missing_entries}")
print(" model syntax and class declarations verified without import")
print(f" registry aliases verified: {len(EXPECTED_REGISTRY_ENTRIES)}")
print("\nDone. Registry aliases installed; do not edit /model/config.json.")
if __name__ == "__main__":
main()