[yangzhuxinyzx]

Purpose

#1277 (DeepSeek-V4.1-Flash) made the pynccl path reduce every all-reduce of 4,194,304 or more elements in place, for all models. vllm.all_reduce is registered as an out-of-place op (direct_register_custom_op without mutates_args, dispatched to _all_reduce_out_place), so compiled graphs may read its input again after the call. Reducing in place hands those readers the summed tensor instead of the rank's own partial.

Qwen3.8-27B NVFP4 + DFlash2 (TP4, hidden 5120) crosses the threshold on every prefill of about 820 tokens or more. The target is not affected, but the DFlash2 draft context is, and acceptance drops.

This PR makes an unset VLLM_INPLACE_AR_MIN mean out-of-place, except under VLLM_V41_PACKED_KV=1. There the DeepSeek-V4.1 packed-KV prefill keeps the in-place reduction it was qualified with (its launcher benchmarks/deepseek_v41/tools/run_serve3.sh already sets VLLM_V41_PACKED_KV=1). An explicit VLLM_INPLACE_AR_MIN value still wins. vllm/envs.py resolves the default, so the communicator only reads envs.VLLM_INPLACE_AR_MIN and gains no env or model coupling.

The second commit also removes the model name that #1265 added to two comments in SpeculativeConfig. That name pushed vllm/config/speculative.py over its check-layering baseline (142 -> 144).

Test Plan

• tests/distributed/test_inplace_all_reduce_default.py (new, CPU): unset → 0, packed KV → 4194304, explicit value wins in both modes.
• tests/test_envs.py (58 passed), tools.generate_env_reference --check, tools/pre_commit/check_env_metadata.py, tools/pre_commit/check_layering.py, ruff 0.14.0 check/format.
• Effective threshold on import: unset 0; VLLM_V41_PACKED_KV=1 4194304; VLLM_V41_PACKED_KV=1 VLLM_INPLACE_AR_MIN=0 0; VLLM_INPLACE_AR_MIN=8388608 8388608.
• End to end on 4x V100-SXM2-16GB (TP4, 300 W, 1530 MHz), original unsloth Qwen3.8-27B NVFP4 + DFlash2 draft7, the shipped profile (262144 max length, E4M3 KV, max-num-seqs 4). Fixed admission fixtures: 8 prompts x 600 tokens, seed 123, temperature 0.7, two repeats in one process.

Test Result

Speed-fixture tokens per round (1K / 8K input). Each row is deterministic across both repeats:

Build	1K	8K
8a3c030 (before #1242)	2.914	2.950
1d7321b (#1242), dbc1f4f (#1252)	2.914	2.950
899e9ed (#1277)	2.581	2.551
c59b3c4 (main)	2.655	2.613
c59b3c4 + VLLM_PYNCCL_ALLGATHER=0	2.655	2.613
c59b3c4 + VLLM_INPLACE_AR_MIN=0	2.875	2.974
this PR, no environment	2.875	2.974

Greedy check, main vs VLLM_INPLACE_AR_MIN=0, 300 tokens on four 1K and four 8K fixtures: identical tokens on all 8. Logprobs are bitwise identical at 1K and differ by at most 0.032 at 8K. So the target output is preserved, and the lost acceptance comes from the draft context.

Short official-style prompts (GSM8K / MT-Bench / HumanEval, 16 each, temperature 1.0, thinking on) never reach the threshold. Their acceptance length was unchanged between 8a3c030 and main (5.31/3.73/4.78 vs 5.24/3.81/4.86). The regression only shows up for long prompts.

check-layering still reports vllm/model_executor/layers/fused_moe/runner/moe_runner.py: platform coupling 37 -> 39. That comes from #1254 (merged 2026-10-24), it is the same on main, and this PR leaves it unchanged. Fixing it needs the round-cost recorder hook moved behind the platform boundary, which belongs in its own change.

Not run: DeepSeek-V4.1 serving. Its behaviour is unchanged by construction under VLLM_V41_PACKED_KV=1. Whether its DSpark context has the same exposure was not checked.

🤖 Generated with Claude Code