DFlash: support vLLM inference backend

[zixi-qi]

Summary

Adds vLLM as a supported inference backend for DFlash draft training. Previously DFlash was gated to inference_engine_type='sgl' only.

• torchspec/train_entry.py — _validate_and_configure_dflash now accepts inference_engine_type in ('vllm', 'sgl') instead of sgl-only, with a clearer error message naming the allowed values.
• configs/vllm_qwen3_8b_dflash.yaml — new DFlash training config for Qwen3-8B on the vLLM backend; the vLLM counterpart to the existing sglang_qwen3_8b_dflash.yaml (same DFlash architecture/hyperparameters, target hidden states generated by vLLM via extract_hidden_states + MooncakeHiddenStatesConnector). 4-GPU layout: 2 inference (tp_size=2) + 2 training (FSDP FULL_SHARD).

Test plan

• ruff check + ruff format --check pass
• pre-commit run --files ... passes (trailing-whitespace, end-of-file-fixer, check-yaml, ruff, ruff-format)
• New config parses as valid YAML
• End-to-end DFlash+vLLM training run on a 4-GPU node

E2E verification

Ran the new config end-to-end on a GB300 node (aarch64, CUDA 13, torch 2.11, vLLM 0.23.0):

./examples/qwen3-8b-single-node/run.sh configs/vllm_qwen3_8b_dflash.yaml

Result: completed cleanly — 3 epochs / 750 steps in 288.7s, checkpoint saved to outputs/qwen3-8b-dflash-vllm/checkpoints/iter_0000751, vLLM engine shut down gracefully.

Metric	Value (final / steady-state)
Loss	0.018
Accuracy	0.992
Accepted length (acc_len)	~14.3 (block_size 16)
Throughput	~4.5 step/s
Per-step compute	fwd ~83ms · bwd ~110ms · opt ~14ms (step ~0.21s)
Inference / training balance	avg 10.4 / 10.4 entries/s, wait=0.0s throughout

The disaggregated pipeline behaved as designed: the vLLM engine (tp_size=2) streamed target hidden states through Mooncake, and the sample-pool backpressure cycled normally (fills to 64/64 → pauses generation → resumes at ~52–56), with the trainer never starving (wait=0.0s).

🤖 Generated with Claude Code