From 5e618bf4806d9266048b36ce7054b7d36ae213d5 Mon Sep 17 00:00:00 2001
From: muh-bot <muh@project6.dev>
Date: Wed, 5 Aug 2026 07:17:37 +0000
Subject: [PATCH] =?UTF-8?q?[housekeeping]=20add=20enginex=5Fbase/=20to=20g?=
 =?UTF-8?q?itignore=20=E2=80=94=20=E8=A7=A3=E5=8E=8B=E5=90=8E=E7=9A=84?=
 =?UTF-8?q?=E5=9F=BA=E7=A1=80=E5=BC=95=E6=93=8E=E4=BB=85=E4=BE=9B=E6=9C=AC?=
 =?UTF-8?q?=E5=9C=B0=E5=88=86=E6=9E=90?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

基础引擎结构 (排除 pkgs 预编译包):
  顶层: Dockerfile, computility-run.yaml, paged_attn.py, prefix_prefill.py, attention.py
  qwen3_6_scripts/: 我们的全部适配代码 (模型定义, attention, API, parsers, patches)
  vllm/: vllm 0.6.3 BI-V100 适配版 (985 files)
    _custom_ops.py: ixformer 绑定层 (V1 decode, silu_and_mul, rms_norm...)
    attention/selector.py: backend 选择 (BI-V100 → XFORMERS → _run_sdpa_fallback)
    model_executor/models/: 200+ 模型定义 (无 qwen3_5)

关键发现:
  1. 基础引擎 Dockerfile 只有 4 行: FROM + COPY + RUN patch_ops.sh
  2. vllm/_custom_ops.py 硬编码 SMEM=32KB (已在我们版本修为 49152)
  3. attention selector: BI-V100 → XFORMERS (not FlashAttn, SM<80)
  4. Prefill 路径: xformers._run_sdpa_fallback (PyTorch matmul+softmax)
     不是 paged_attn.forward_prefix (那是 context attention for cached tokens)

CCCL source read: dispatch_batch_memcpy.cuh (large_buffer vs small_buffer paths)
  KV block size = 16×4×256×2 = 32KB → large_buffer path in CCCL
---
 .gitignore | 1 +
 1 file changed, 1 insertion(+)

diff --git a/.gitignore b/.gitignore
index 8a60547a..4a39145b 100644
--- a/.gitignore
+++ b/.gitignore
@@ -4,3 +4,4 @@ cccl_upstream/
 muh/
 baseline.muh
 pkgs/
+enginex_base/