8030a11b96
project_7 docker build 已在竞赛平台验证成功。 完整搬运 wudixzy/competition stack: - qwen3_5.py 2615 行 (12 个 corex .so 调用) - patch_ops.sh 251 行 (set -eo pipefail + cd dirname) - 12 prebuilt corex .so (SHA256 verified) - 13 CUDA .cu 源码 + 11 build scripts - 9 vendor overrides (block/sampler/scheduler) - transformers-4.55.3 offline wheel - computility-run.yaml: 262144 max-model-len, BI100 env vars - Dockerfile 结构不变 (COPY qwen3_6_scripts + RUN patch_ops.sh)
82 lines
3.4 KiB
Python
82 lines
3.4 KiB
Python
from patch_utils import package_root, replace_one_of
|
|
|
|
WORKER = package_root("vllm") / "worker" / "worker.py"
|
|
|
|
CLEAN_BLOCK = """\
|
|
# Profile the memory usage of the model and get the maximum number of
|
|
# cache blocks that can be allocated with the remaining free memory.
|
|
torch.cuda.empty_cache()
|
|
|
|
# Execute a forward pass with dummy inputs to profile the memory usage
|
|
# of the model.
|
|
self.model_runner.profile_run()
|
|
"""
|
|
|
|
GUARDED_BLOCK = """\
|
|
# Profile the memory usage of the model and get the maximum number of
|
|
# cache blocks that can be allocated with the remaining free memory.
|
|
torch.cuda.empty_cache()
|
|
|
|
# Execute a forward pass with dummy inputs to profile the memory usage
|
|
# of the model. Mark this synthetic pass so BI100_PROFILE can skip
|
|
# timing it by default; profiling real requests is the useful signal.
|
|
_bi100_prev_startup_profile = os.environ.get("BI100_IN_STARTUP_PROFILE")
|
|
os.environ["BI100_IN_STARTUP_PROFILE"] = "1"
|
|
try:
|
|
self.model_runner.profile_run()
|
|
finally:
|
|
if _bi100_prev_startup_profile is None:
|
|
os.environ.pop("BI100_IN_STARTUP_PROFILE", None)
|
|
else:
|
|
os.environ["BI100_IN_STARTUP_PROFILE"] = _bi100_prev_startup_profile
|
|
"""
|
|
|
|
NEW_BLOCK = """\
|
|
# Profile the memory usage of the model and get the maximum number of
|
|
# cache blocks that can be allocated with the remaining free memory.
|
|
torch.cuda.empty_cache()
|
|
|
|
# BI100: Qwen3.6 batched dummy profile_run can trip GDN non-finite
|
|
# checks before the server starts. If the operator explicitly provides
|
|
# --num-gpu-blocks-override, trust that conservative capacity value and
|
|
# skip only the synthetic profile pass. Real inference still uses the
|
|
# normal GDN fail-fast path.
|
|
if self.cache_config.num_gpu_blocks_override is not None:
|
|
cache_block_size = self.get_cache_block_size_bytes()
|
|
if cache_block_size == 0:
|
|
num_cpu_blocks = 0
|
|
else:
|
|
num_cpu_blocks = int(self.cache_config.swap_space_bytes //
|
|
cache_block_size)
|
|
logger.warning(
|
|
"[BI100] skipping worker.profile_run because "
|
|
"num_gpu_blocks_override=%d was explicitly set",
|
|
self.cache_config.num_gpu_blocks_override)
|
|
gc.collect()
|
|
torch.cuda.empty_cache()
|
|
return self.cache_config.num_gpu_blocks_override, max(num_cpu_blocks, 0)
|
|
|
|
# Execute a forward pass with dummy inputs to profile the memory usage
|
|
# of the model. Mark this synthetic pass so BI100_PROFILE can skip
|
|
# timing it by default; profiling real requests is the useful signal.
|
|
_bi100_prev_startup_profile = os.environ.get("BI100_IN_STARTUP_PROFILE")
|
|
os.environ["BI100_IN_STARTUP_PROFILE"] = "1"
|
|
try:
|
|
self.model_runner.profile_run()
|
|
finally:
|
|
if _bi100_prev_startup_profile is None:
|
|
os.environ.pop("BI100_IN_STARTUP_PROFILE", None)
|
|
else:
|
|
os.environ["BI100_IN_STARTUP_PROFILE"] = _bi100_prev_startup_profile
|
|
"""
|
|
|
|
replace_one_of(
|
|
WORKER,
|
|
[
|
|
(GUARDED_BLOCK, NEW_BLOCK),
|
|
(CLEAN_BLOCK, NEW_BLOCK),
|
|
],
|
|
required=True,
|
|
already_contains="[BI100] skipping worker.profile_run",
|
|
)
|