test: 回退Docker context到26e6cb40完全一致——验证竞赛平台build

Dockerfile/qwen3_6_scripts/ex_engine/computility-run.yaml 全部
还原到26e6cb40的精确内容。删除所有26e6cb40不存在的新增文件
(prebuilt/*.so, wheels/*.whl, vendor_overrides/, 新增.cu/.sh等)。

目的:确认26e6cb40的文件内容在当前git状态下仍能通过竞赛平台build。
如果通过,说明问题在新增文件中;如果不通过,说明问题在git仓库层面。
This commit is contained in:
Claude
2026-08-11 18:06:09 +00:00
parent af2258f32a
commit 6f6b7e959b
135 changed files with 10134 additions and 31627 deletions

View File

@@ -1,121 +0,0 @@
"""Install disabled-by-default M1-48 XFormers timing boundaries."""
from __future__ import annotations
from pathlib import Path
try:
from patch_utils import package_root, replace_once
except ModuleNotFoundError:
from .patch_utils import package_root, replace_once
IMPORT_OLD = "from vllm.logger import init_logger"
IMPORT_NEW = """\
from vllm.bi100_profile import bi100_timer
from vllm.logger import init_logger"""
KV_WRITE_OLD = """\
PagedAttention.write_to_paged_cache(key, value, key_cache,
value_cache,
updated_slot_mapping,
self.kv_cache_dtype,
k_scale, v_scale)"""
KV_WRITE_NEW = """\
with bi100_timer("xformers.kv_write"):
PagedAttention.write_to_paged_cache(
key, value, key_cache, value_cache,
updated_slot_mapping, self.kv_cache_dtype,
k_scale, v_scale)"""
DENSE_OLD = """\
out = self._run_memory_efficient_xformers_forward(
query, key, value, prefill_meta, attn_type=attn_type)"""
DENSE_NEW = """\
with bi100_timer("xformers.dense_prefill"):
out = self._run_memory_efficient_xformers_forward(
query, key, value, prefill_meta, attn_type=attn_type)"""
PAGED_OLD = """\
out = PagedAttention.forward_prefix(
query,
key,
value,
self.kv_cache_dtype,
key_cache,
value_cache,
prefill_meta.block_tables,
prefill_meta.query_start_loc,
prefill_meta.seq_lens_tensor,
prefill_meta.context_lens_tensor,
prefill_meta.max_query_len,
self.alibi_slopes,
self.sliding_window,
k_scale,
v_scale,
is_causal_decoder=(attn_type == AttentionType.DECODER),
)"""
PAGED_NEW = """\
with bi100_timer("xformers.paged_prefill"):
out = PagedAttention.forward_prefix(
query,
key,
value,
self.kv_cache_dtype,
key_cache,
value_cache,
prefill_meta.block_tables,
prefill_meta.query_start_loc,
prefill_meta.seq_lens_tensor,
prefill_meta.context_lens_tensor,
prefill_meta.max_query_len,
self.alibi_slopes,
self.sliding_window,
k_scale,
v_scale,
is_causal_decoder=(attn_type == AttentionType.DECODER),
)"""
def patch_file(path: Path) -> None:
replace_once(
path,
IMPORT_OLD,
IMPORT_NEW,
already_contains="from vllm.bi100_profile import bi100_timer",
)
replace_once(
path,
KV_WRITE_OLD,
KV_WRITE_NEW,
already_contains='bi100_timer("xformers.kv_write")',
)
replace_once(
path,
DENSE_OLD,
DENSE_NEW,
already_contains='bi100_timer("xformers.dense_prefill")',
)
replace_once(
path,
PAGED_OLD,
PAGED_NEW,
already_contains='bi100_timer("xformers.paged_prefill")',
)
text = path.read_text(encoding="utf-8")
canonical = "\n".join(line.rstrip(" \t") for line in text.split("\n"))
if not canonical.endswith("\n"):
canonical += "\n"
if canonical != text:
path.write_text(canonical, encoding="utf-8")
def main() -> None:
path = package_root("vllm") / "attention" / "backends" / "xformers.py"
print("=== patch_xformers_profile (M1-48 diagnostic timers) ===")
print(f"Target: {path}")
patch_file(path)
if __name__ == "__main__":
main()