[Experiment][SM70] Qualify E4M3 KV with MTP3

[Leonccaa]

Problem and resulting behavior

Qwen4Exp QSA rejected every E4M3 target KV configuration that also had a speculative drafter, even though the check recorded the original MTP0 qualification boundary. This draft removes that admission-only check and records the additional runtime work required for a safe MTP3 configuration.

The current branch now:

• permits the combined configuration to reach real SM70 qualification;
• finalizes QSA scale state for standalone MTP loading and refuses an E4M3 draft whose calibrated K/V scales are missing;
• allocates the Flash-V100 XQA temporary output in FP32 for E4M3 KV, matching the native ABI;
• honors speculative_config.kv_cache_dtype in the Model Runner V2 Eagle/MTP loader.

The SM70, FP16 activation, TP4, cache-format, and calibrated target-scale gates remain intact.

Integration baseline

• Aligns the fork with 1CatAI/main@02c87ab890, including [Kernel][SM70] Share FP16 attention fast paths with E4M3 KV 1CatAI/1Cat-vLLM#638.
• Preserves the deployed CPU-cache H2D ordering fix.
• Includes the current [Bugfix][Core] Backport upstream Mamba prefix-cache retention 1CatAI/1Cat-vLLM#617 and [CI] Fix baseline QSA typing and pre-commit failures 1CatAI/1Cat-vLLM#624 changes already present in the fork baseline.
• Includes the deployed [Bugfix][MRV2] Backport phase-aware decode graph dispatch 1CatAI/1Cat-vLLM#637 phase-aware MRV2 dispatch fix.
• Completes the current [KV Offload] Grouped hybrid RAM caching, filesystem restoration and retention-sized Mamba pools (stacked on #617) 1CatAI/1Cat-vLLM#598 stack with its final documentation and Mamba boundary hand-off fix.

Duplicate-work check

Searched both 1CatAI/1Cat-vLLM and Leonccaa/1Cat-vLLM for E4M3, MTP, and QSA work. 1CatAI#639 admits checkpoint-native FP8 MTP expert weights under pipeline parallelism; it does not change target KV dtype or the QSA E4M3/MTP0 admission check.

Validation

Static and CPU checks

• Changed-file pre-commit passed after each change, including Ruff, format, mypy, SPDX, configuration validation, and repository checks.
• Initial focused CPU test bodies: 17 passed and 7 CUDA cases skipped. The no-GPU host reported 16 teardown-only errors when the shared cleanup fixture called the accelerator API.
• Draft scale/XQA workspace tests: 2 passed in the release container.
• MRV2 draft KV override and Eagle weight-sharing tests: 8 passed in the release container.
• Full native CPU build completed without OOM; 15 native modules and all manifest-tracked Python files passed the release-image hash/import verifier.

CT252 runtime findings

The runtime used the existing Qwen3.8 Flash-Next AWQ g32 target, native FP8 MTP3 weights, TP4, 262,144 max context, 8,192 batched-token budget, C4, and 32 GiB CPU KV offload.

1. Target and draft both inherited E4M3: the service started, loaded all 24/24 calibrated target K/V scales, captured the C4/mixed grouped-XQA graphs, and allocated 867,200 GPU KV tokens (3.31 full 262,144-token requests). The draft checkpoint had 0/2 QSA K/V scales. Runtime acceptance failed: the [Bugfix][MRV2] Backport phase-aware decode graph dispatch 1CatAI/1Cat-vLLM#637 804-token regression returned duct plus repeated punctuation instead of OK; all four short concurrent responses were wrong or empty; forced tool choice returned HTTP 500 after the grammar rejected token 0; speculative acceptance was 0/510 drafted tokens.
2. E4M3 target with explicit FP16 draft KV: after fixing MRV2 to honor the draft override, the scale failure disappeared and target/draft weights loaded. KV planning then failed with CSA+linear layer 48 violates cache geometry, showing that the current cache grouper cannot build this mixed target/draft cache layout.
3. Rollback control: the preserved [Bugfix][MRV2] Backport phase-aware decode graph dispatch 1CatAI/1Cat-vLLM#637 FP16 target-KV/MTP3/C2 instance returned the same 804-token probe as exact OK in 0.58 seconds. This distinguishes the new E4M3/MTP interaction from the original [Bugfix][MRV2] Backport phase-aware decode graph dispatch 1CatAI/1Cat-vLLM#637 K+1 prefill/decode classification bug.

Status

This PR remains draft. It is useful as the integration and evidence branch, but neither tested draft-cache arrangement is production-safe yet. The next gate is either calibrated E4M3 scales for the MTP QSA layer or cache grouping support for an FP16 draft beside an E4M3 target, followed by the same 804-token, tool-call, C4, and business-workload acceptance.

AI assistance from OpenAI Codex was used to inspect the dependency stack, prepare the integration branch, implement the changes, and run validation. Leon requested the configuration experiment and will review the result before production acceptance.