Qwen3.6-27B-AEON-Ultimate-Uncensored-DFlash/Dockerfile.v3 at main · AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-DFlash · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
# vllm-aeon-ultimate-dflash:qwen36-v3 — 2026-04-28 fresh build
# Source-built vLLM v0.20.0 release commit (88d34c6409) targeting GB10 / sm_121a / DGX Spark
#
# What's NEW in v3 vs v2.1:
#   ✓ vLLM commit 88d34c6409 — actual v0.20.0 release commit (published 2026-04-27)
#       vs v2.1's 101584af0a (cut 2026-04-23, before release was finalized).
#       Brings ~4 days of v0.20.0-branch bugfixes + the official release seal.
#   ✓ FlashInfer v0.6.9 stable (released 2026-04-24) — replaces v0.6.9rc1 (2026-04-23).
#       Same b12x SM121 NVFP4 GEMM backend, with rc-cycle fixes folded in.
#   ✓ Same patch surface as v2.1 (verified upstream still missing the relevant
#       fixes as of 2026-04-28 — see audit notes per patch below).
#
# What stays the same as v2.1 (deliberately — to isolate the vLLM/FlashInfer
# upgrade as the only variable):
#   • Base image: ghcr.io/aeon-7/vllm-spark-gemma4-nvfp4-awq:latest (CUDA 13.2 + torch nightly cu130)
#   • TORCH_CUDA_ARCH_LIST="12.0+PTX" (JITs to sm_121a at runtime; safer than pinning)
#   • TurboQuant @ AEON-7 fork (CUDA-graph-safe _POWERS cache; until 0xSero#12 merges)
#   • Speculators v0.3 still SKIPPED (pydantic incompat unfixed in v0.3.0; vLLM has
#       native --speculative-config so we don't actually need Speculators)
#   • All 5 .py patches (idempotent — print "already applied" if upstream landed it)
#
# Audit per patch as of 2026-04-28 against vLLM v0.20.0 release tag:
#   register_qwen3_5_text.py    — STILL NEEDED. PRs #36289/#36607/#36850 still closed
#                                 unmerged; Qwen3_5MoeForCausalLM not in upstream registry.
#   patch_cuda_optional_import.py — STILL NEEDED. RTLD_LAZY workaround for SM121 builds
#                                 missing SM100-only mxfp4_experts_quant kernels.
#                                 ENABLE_NVFP4_SM100=0 env helps at build time but
#                                 dlopen at runtime still resolves all symbols by
#                                 default; need lazy mode.
#   patch_kv_cache_utils.py     — STILL NEEDED. v0.20.0 still has unfiltered
#                                 `min_block_size = min([... for group in groups])`
#                                 which crashes on hybrid linear-attn groups (block_size=None).
#   patch_mrope_text_fallback.py — STILL NEEDED. Qwen3_5 doesn't implement
#                                 SupportsMRoPE in vLLM HEAD; we provide the
#                                 text-only T=H=W=arange fallback that matches
#                                 the DFlash drafter's standard 1D RoPE distribution.
#   patch_cudagraph_align.py    — STILL NEEDED defensively. PRs #40092/#40454 in
#                                 v0.20.0 fixed *some* of the spec-decode capture
#                                 alignment issues, but the FULL-only gate on the
#                                 alignment filter in compilation.py persists, so
#                                 PIECEWISE mode still silently skips alignment.
#                                 Patch is idempotent — will no-op if upstream caught up.
#
# Build time on Spark: 25-50 min on warm ccache (post v2.1 build), 60-90 min cold.
# Build:  docker build -t ghcr.io/aeon-7/vllm-aeon-ultimate-dflash:qwen36-v3 -f Dockerfile.v3 .
# Push:   docker push ghcr.io/aeon-7/vllm-aeon-ultimate-dflash:qwen36-v3

FROM ghcr.io/aeon-7/vllm-spark-gemma4-nvfp4-awq:latest

RUN apt-get update && apt-get install -y --no-install-recommends \
      ccache git \
 && rm -rf /var/lib/apt/lists/*

ENV PIP_NO_CACHE_DIR=1 \
    UV_SYSTEM_PYTHON=1 \
    TORCH_CUDA_ARCH_LIST="12.0+PTX" \
    MAX_JOBS=14 \
    NVCC_THREADS=2 \
    CMAKE_BUILD_PARALLEL_LEVEL=14 \
    CCACHE_DIR=/root/.ccache \
    USE_CCACHE=1 \
    CUDA_HOME=/usr/local/cuda \
    PATH=/usr/local/cuda/bin:$PATH \
    VLLM_TEST_FORCE_FP8_MARLIN=1 \
    ENABLE_NVFP4_SM100=0

RUN python3 -c "import torch; print(f'torch={torch.__version__} CUDA={torch.version.cuda}')" && \
    nvcc --version | tail -1 && \
    ccache --version | head -1

# Pin to v0.20.0 RELEASE commit (the actual published-release ref, 2026-04-27).
# Our v2.1 used 101584af0a (cut 2026-04-23 from the v0.20.0 branch but pre-release).
ARG VLLM_COMMIT=88d34c6409e9fb3c7b8ca0c04756f061d2099eb1
RUN git clone https://github.com/vllm-project/vllm.git /workspace/vllm-src \
 && cd /workspace/vllm-src \
 && git checkout ${VLLM_COMMIT} \
 && git log -1 --format='vLLM build pin: %H %s'

WORKDIR /workspace/vllm-src

RUN python3 use_existing_torch.py

RUN uv pip install --system -r requirements/build.txt 2>/dev/null || \
    uv pip install --system -r requirements/build/cuda.txt

RUN uv pip install --system --no-build-isolation --no-deps . 2>&1 | tee /tmp/vllm-build.log | tail -100

# FlashInfer v0.6.9 STABLE (2026-04-24) — replaces v0.6.9rc1 from v2.1.
RUN uv pip uninstall --system flashinfer-python flashinfer-cubin flashinfer-jit-cache 2>/dev/null || true
RUN uv pip install --system --no-deps \
      "flashinfer-python @ git+https://github.com/flashinfer-ai/flashinfer.git@v0.6.9"

RUN uv pip install --system "scipy>=1.11"
RUN uv pip install --system --no-deps \
      "turboquant @ git+https://github.com/AEON-7/turboquant.git@fix/cuda-graph-safe-qjl-powers"

RUN python3 -c "import transformers; print(f'pre-pin transformers={transformers.__version__}')" && \
    uv pip install --system "transformers>=5.6,<6"

WORKDIR /
RUN rm -rf /workspace/vllm-src

# === Patches (audit notes in header; all idempotent) ===

COPY patches/register_qwen3_5_text.py /opt/patches/
RUN python3 /opt/patches/register_qwen3_5_text.py

COPY patches/patch_cuda_optional_import.py /opt/patches/
RUN python3 /opt/patches/patch_cuda_optional_import.py

COPY patches/patch_kv_cache_utils.py /opt/patches/
RUN python3 /opt/patches/patch_kv_cache_utils.py

COPY patches/patch_mrope_text_fallback.py /opt/patches/
RUN python3 /opt/patches/patch_mrope_text_fallback.py

COPY patches/patch_cudagraph_align.py /opt/patches/
RUN python3 /opt/patches/patch_cudagraph_align.py

# === Verification ===
RUN python3 -c "\
import torch, vllm, flashinfer, turboquant; \
print(f'POST vLLM={vllm.__version__}'); \
print(f'POST flashinfer={flashinfer.__version__}'); \
print(f'POST torch={torch.__version__} CUDA={torch.version.cuda}'); \
print(f'POST turboquant={getattr(turboquant, \"__version__\", \"?\")}'); \
from vllm.model_executor.models.registry import _TEXT_GENERATION_MODELS as T; \
assert 'Qwen3_5MoeForCausalLM' in T, 'registry patch not applied'; \
print('Qwen3_5MoeForCausalLM ->', T['Qwen3_5MoeForCausalLM']); \
" && \
    ls /usr/local/lib/python3.12/dist-packages/vllm/_C.abi3.so && \
    echo 'image OK — vllm._C will load at runtime when --gpus all mounts libcuda.so.1'

LABEL org.opencontainers.image.title="vllm-aeon-ultimate-dflash" \
      org.opencontainers.image.description="AEON-7 vLLM container for Qwen3.6-27B-AEON-Ultimate-Uncensored on DGX Spark / GB10. v3: vLLM v0.20.0 release commit (88d34c6409, 2026-04-27) + FlashInfer v0.6.9 stable + 5 sm_121a patches (idempotent). DFlash speculative decoding (k=15) via z-lab/Qwen3.6-27B-DFlash drafter." \
      org.opencontainers.image.source="https://github.com/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-DFlash" \
      org.opencontainers.image.url="https://github.com/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-DFlash" \
      org.opencontainers.image.documentation="https://github.com/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-DFlash#readme" \
      org.opencontainers.image.version="v3" \
      org.opencontainers.image.base.name="ghcr.io/aeon-7/vllm-spark-gemma4-nvfp4-awq:latest" \
      vllm.compute_capability="sm_120+PTX" \
      vllm.target_hardware="DGX Spark / GB10 / sm_121a" \
      vllm.vllm_commit="88d34c6409e9fb3c7b8ca0c04756f061d2099eb1"