[Misc][DSv4] Official Spark GB10 container on vLLM v0.28.0

[asterayx]

Purpose

Official DGX Spark / GB10 container line for DeepSeek-V4-Flash, rooted at official vLLM v0.28.0 (2cf0a6915), plus only the Spark/SM12x runtime patches that made dual-node serve green.

This is a replacement line, not a delta on today’s asterayx/vllm main. Base is the v0.28.0 branch (the official tag).

Latest: Vision-Exp mixed-warmup _out (285a71df4)

Worker died after SM12x: insert 4 extra KV slots for prefill pad 2 -> 6:

AttributeError: '_OpNamespace' '_C' object has no attribute 'fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert_out'

v0.28.0 eager_scratch_pool calls that _out op. later-main never allocated the pool and used allocating fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert. Host-mounted later-main _C_stable_libtorch does not register _out. SM12x now skips the pool and getattr-guards _out. No CUDA rebuild.

Previous: NameError: copy in DSpark capture (68c95e735) is fixed.

vs later-main (cursor/dsv4-flash-vision-exp-df88)

Vision-Exp serve path (k=3, in-vocab sentinels, vision gate, DSpark draft, Eagle3, FlashInfer dual-cache, graph-safe VL topk) was already on this line.

#52018 leftovers that were still missing and are now ported (285580c8a):

• NVFP4 VLLM_B12X_MOE_FP4_FORCE_A16 / use_a16 through select, convert, quant config, ModelOpt, compressed_tensors
• Provider-based MoE warmup (linear warmup kept for v0.28.0 split kernels)
• Eager b12x submodule import ([Bugfix] Resolve B12X modules before Dynamo tracing vllm-project/vllm#53326 Dynamo)
• Docs, test_b12x.py, nvfp4_utils layout flags, INC packing parametrize

Intentionally not copied from later-main (v0.28.0 architecture / dropped on purpose):

• later-main multimodal processor.py rewrite (v0.28.0 still pops input_ids; we tokenize on the VL processor instead)
• later-main unified scaled_mm/b12x.py / linear get_b12x_warmup_unit ([Refactor] Simplify B12X linear kernels and warmup vllm-project/vllm#52368)
• Buildkite DGX Spark nightly job
• FlashInfer mega-MoE EP backends
• debug.sh / gdb-dump.sh

AI assistance was used (Cursor Grok 4.6).

Test Plan

Cloud (x86 CPU, no GPU):

.venv/bin/python -m pytest \
  tests/utils/test_sm12x.py::test_sm12x_align_decode_q_len_snaps_to_safe_widths \
  tests/utils/test_sm12x.py::test_sm12x_align_tokens_unchanged_off_sm12x \
  tests/utils/test_sm12x.py::test_fused_qnorm_insert_out_is_optional \
  tests/utils/test_sm12x.py::test_dspark_init_cudagraph_manager_copies_capture_sizes \
  -v --noconftest
# 4 passed

Spark (host-mount Vision-Exp; worker then head):

git pull origin cursor/spark-v0280-dsv4-df88
NODE_RANK=1 VLLM_HOST_IP=192.168.100.11 HEADLESS=--headless ./docker/gb10/run-vision.sh
NODE_RANK=0 ./docker/gb10/run-vision.sh

Look for mixed warmup after SM12x: insert 4 extra KV slots for prefill pad 2 -> 6, not fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert_out AttributeError.

Why this is not a duplicate

Open Spark PRs on later-main (#5 observability, #7 FlashInfer 0.6.18) are a different history. #52018 is not an ancestor of v0.28.0.