Honor the DFlash draft load config

[FujitsuPolycom]

Resulting behavior

The composed vLLM 0b67266a0f37d6146a8403fb8482403c62f412d5 runtime passes SpeculativeConfig.draft_load_config into the DFlash get_model() call. A profile can therefore keep global target loading on fastsafetensors while loading the external BF16 DFlash2 checkpoint with standard safetensors.

The exact configuration shape is:

--load-format fastsafetensors
--speculative-config '{"method":"dflash","model":"/mtp-draft","num_speculative_tokens":7,"draft_tensor_parallel_size":4,"kv_cache_dtype":"auto","draft_sample_method":"probabilistic","rejection_sample_method":"standard","draft_load_config":{"load_format":"safetensors"}}'

draft_load_config remains optional. When it is absent or null, vLLM's existing get_model(load_config or vllm_config.load_config) rule retains the global target loader for the draft.

Exact source contract

• vLLM commit: 0b67266a0f37d6146a8403fb8482403c62f412d5
• vLLM tree: ba9484ccb33aa56e90ff2f447f15ca9b9da97639
• target: vllm/v1/worker/gpu/spec_decode/dflash/utils.py
• preimage SHA-256: 2301c8199b73ed893dfbd3ae14ad125816f100b2d2ed034215b1f2d9aa2c23c5
• patch SHA-256: 39b567013ee7aed79f63200ed460129587933dc77fb430decdf19f78178de279
• postimage SHA-256: 98acbae2b3bb4482d83f9637c163ce7c92707ccdf6561b7e431f23337f151cf4

The source preparer applies and reverses the patch against the immutable vLLM checkout to prove both byte boundaries. The prepared-context receipt, image labels, composed verifier, and profile label requirements bind the patch and postimage. GPU-free AST checks reject a DFlash loader that ignores draft_load_config and reject a model-loader implementation that loses the None fallback.

Compatibility

This draft was created from exact SparkRing #131 commit a7e5a70af1b7cb329651712a449c60f54b1d0ae5. The #131 branch later added unrelated README commit 56b3a8f77d5b35a63bb97c213a73d34e5c178db6; that commit is deliberately not folded into this exact-base branch.

No vLLM, B12X, SparkCache, NCCL, model, or image source commit changed. Existing MTP profile identity fields and SparkCache cache identity are unchanged. Cache namespace impact: none.

No image was built or published. No service or serving host was contacted or changed.

Validation

• Focused runtime, resolver, terminology, and launcher tests: 42 passed.
• DFlash runtime and resolver tests: 23 passed.
• Fresh source-context preparation followed by independent receipt verification: passed.
• Ruff, JSON parsing, Python compilation, Bash syntax, and diff checks: passed.
• Maintained exact-base suite: 1,935 passed, 9 skipped, 1 inherited README assertion failure. The later Use SparkCache CUDA restore terminology #131 commit named above addresses that unrelated assertion outside this requested base.

[FujitsuPolycom]

The resulting runtime, operator contract, and evidence are consolidated in retained draft stack #146 → #147 → #150. Independent page-base research remains in #149. Closing this superseded draft and deleting only its remote head branch.