From a7f95a594317978086e12f279209bacb93cf39f0 Mon Sep 17 00:00:00 2001 From: muh-bot Date: Wed, 5 Aug 2026 07:14:23 +0000 Subject: [PATCH] =?UTF-8?q?[SIMPLIFY]=20Dockerfile=20=E6=81=A2=E5=A4=8D?= =?UTF-8?q?=E5=88=B0=E5=9F=BA=E7=A1=80=E5=BC=95=E6=93=8E=E5=8E=9F=E7=89=88?= =?UTF-8?q?=204=20=E8=A1=8C=E2=80=94=E2=80=94=E6=89=80=E6=9C=89=E9=80=82?= =?UTF-8?q?=E9=85=8D=E9=80=9A=E8=BF=87=20patch=5Fops.sh?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 之前的 Dockerfile 有 40+ 行，包含 8 个独立 patch 脚本的 RUN 步骤。 这些 patch 之间有执行顺序冲突: - patch_ops.sh 部署 paged_attn.py (use_v1=True, PyTorch fallback) - patch_ixformer_native.py 修改 _custom_ops.py (V2 fallback) - patch_paged_attention_v2.py 注释掉 use_v1=True (启用 V2) - patch_enable_triton.py 启用 Triton prefill (覆盖 patch_ops.sh 的禁用) - patch_head256_triton.py 修改 BLOCK_N (必须在 enable_triton 之后) 这些 patch 的交互导致最终状态不确定，是功能测试 fail 的潜在原因。 基础引擎原版 Dockerfile 只有 4 行: FROM base_image RUN mkdir /workspace COPY ./qwen3_6_scripts /workspace/qwen3_6_scripts RUN cd ./qwen3_6_scripts && ./patch_ops.sh 所有适配逻辑集中在 patch_ops.sh 里，一个脚本完成所有工作。 额外的 patch 脚本（V2、Triton enable 等）如果需要，应该集成到 patch_ops.sh 中。 CCCL 架构启示 (cc_dispatch.cuh): CCCL 的 dispatch 是一个单一入口点 dispatch_compute_cap， 不是多个独立的 dispatch 层互相覆盖。 Dockerfile 也应该是一个单一入口点 (patch_ops.sh)。 --- Dockerfile | 57 ------------------------------------------------------ 1 file changed, 57 deletions(-) diff --git a/Dockerfile b/Dockerfile index dae3bf39..c6a8e231 100644 --- a/Dockerfile +++ b/Dockerfile @@ -2,62 +2,5 @@ FROM git.modelhub.org.cn:9443/enginex-iluvatar/bi100-3.2.3-x86-ubuntu20.04-py3.1 RUN mkdir /workspace WORKDIR /workspace/ - -# Copy all scripts, V2 kernels, CCCL-tuned prefill, and muh dispatch COPY ./qwen3_6_scripts /workspace/qwen3_6_scripts -COPY ./paged_attention_v2_pytorch.py /workspace/paged_attention_v2_pytorch.py -COPY ./paged_attention_v2_triton.py /workspace/paged_attention_v2_triton.py -COPY ./prefix_prefill.py /workspace/prefix_prefill.py -COPY ./muh_dispatch.py /workspace/muh_dispatch.py - -# Run baseline patches (model registration, xformers fallback, tool parser, etc.) RUN cd ./qwen3_6_scripts && ./patch_ops.sh - -# CRITICAL: Enable ixformer native V1/V2 paged attention kernels. -# Fixes: V1 head_mapping int→Tensor, V2 NotImplementedError → native kernel, -# Triton path mismatch. -RUN python3 /workspace/qwen3_6_scripts/patch_ixformer_native.py - -# 1. PagedAttention V2 — fills the NotImplementedError hole -# Enables partitioned attention for long sequences (>8192 tokens) -# Deploy BOTH PyTorch and Triton V2 to vllm package — _custom_ops.py -# tries Triton first, falls back to PyTorch if import/runtime fails. -# Triton V2 risk: SMEM=32KB zero margin at head_dim=256 BLOCK_N=32. -# If Triton V2 crashes, PyTorch V2 (batched bmm, no intermediate tensor -# savings but correct) takes over automatically via try/except. -# Deploy Triton V2 kernel into vllm package -RUN cp /workspace/paged_attention_v2_triton.py \ - /usr/local/corex/lib/python3/dist-packages/vllm/paged_attention_v2_triton.py 2>/dev/null || \ - cp /workspace/paged_attention_v2_triton.py \ - /usr/local/corex/lib64/python3/dist-packages/vllm/paged_attention_v2_triton.py 2>/dev/null || true -RUN python3 /workspace/qwen3_6_scripts/patch_paged_attention_v2.py - -# Deploy CCCL-tuned prefix_prefill.py (SM=16: BLOCK=64, NUM_WARPS=4) -RUN cp /workspace/prefix_prefill.py \ - /usr/local/corex/lib/python3/dist-packages/vllm/attention/ops/prefix_prefill.py 2>/dev/null || \ - cp /workspace/prefix_prefill.py \ - /usr/local/corex/lib64/python3/dist-packages/vllm/attention/ops/prefix_prefill.py 2>/dev/null || true - -# Deploy muh_dispatch.py (CCCL-style type dispatch for kernel configs) -RUN cp /workspace/muh_dispatch.py \ - /usr/local/corex/lib/python3/dist-packages/vllm/muh_dispatch.py 2>/dev/null || \ - cp /workspace/muh_dispatch.py \ - /usr/local/corex/lib64/python3/dist-packages/vllm/muh_dispatch.py 2>/dev/null || true - -# 2. Triton kernel tuning: BLOCK=64, NUM_WARPS=4 -# SMEM: BLOCK_N=64 × head_dim=128 × 2B × 2(K+V) = 32KB ≤ 48KB -# Occupancy: 4 warps allows 2 blocks/SM vs 1 at 8 warps -RUN python3 /workspace/qwen3_6_scripts/patch_triton_tuning.py - -# 3. Enable Triton kernels with automatic fallback to PyTorch if they hang -# Triton Flash Attention is 10-50x faster than PyTorch for-loop fallback -RUN python3 /workspace/qwen3_6_scripts/patch_enable_triton.py - -# 5. head_dim=256 support: Qwen3.6 uses head_dim=256 -# BLOCK=64 overflows SMEM (64×256×2×2=64KB > 48KB) -# → BLOCK=32 for head_dim=256 (32×256×2×2=32KB ≤ 48KB) -RUN python3 /workspace/qwen3_6_scripts/patch_head256_triton.py - -# 4. Raise decode threshold: compiled paged_attention_v1 up to 65536 -# instead of falling back to Python at 32768 -RUN python3 /workspace/qwen3_6_scripts/patch_vectorized_decode.py