{
  "schema_version": "operator-workflow/v1",
  "request": {
    "action": "independent pinned-source review of V2 MTP3 with Triton attention and dynamic FP8 KV",
    "model": "bottlecapai/ThinkingCap-Qwen3.8-27B-NVFP4A4-AWQ",
    "hardware": "RTX 5090 (sm120)",
    "execution_scope": "source review and candidate instructions only; no endpoint, GPU, lifecycle, route, or promotion action"
  },
  "gate_state": "not_required",
  "targets": [
    "configs/thinkingcap-awq-vllm029-5090-16k-mtp3-triton-bf16.toml",
    "configs/thinkingcap-awq-vllm029-5090-16k-mtp3-triton-fp8.toml"
  ],
  "tools_used": [
    "read-only inspection of the exact pinned upstream vLLM revision",
    "read-only review of retained, sanitized campaign evidence"
  ],
  "artifacts": {
    "review_independence": {
      "implementing_model": "GPT-6 Astra",
      "reviewer_model": "GPT-5.6 Sol",
      "evaluated_model": "ThinkingCap Qwen3.8 27B",
      "independent": true
    },
    "pinned_runtime": {
      "vllm_version": "0.29.0",
      "vllm_revision": "98dff2a81d747d1dba01a47f939f48c3526d4206",
      "source_inspected_on": "2026-09-23",
      "source_urls": [
        {
          "url": "https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/config/vllm.py#L2349-L2415",
          "purpose": "V2 model-runner feature gate explicitly admits MTP"
        },
        {
          "url": "https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/config/speculative.py#L385-L391",
          "purpose": "declared draft attention-backend and KV-dtype fields, including target-KV inheritance when the draft KV dtype is unset"
        },
        {
          "url": "https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/worker/gpu/spec_decode/mtp/speculator.py#L11-L33",
          "purpose": "standard MTP loads through load_eagle_model"
        },
        {
          "url": "https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/worker/gpu/spec_decode/eagle/utils.py#L33-L49",
          "purpose": "pinned standard MTP/EAGLE loader does not apply speculative_config.attention_backend and only replaces cache dtype when an explicit draft override is present"
        },
        {
          "url": "https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/worker/gpu/spec_decode/speculator.py#L206-L225",
          "purpose": "standard draft attention metadata builders use the target VllmConfig"
        },
        {
          "url": "https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/worker/gpu/spec_decode/autoregressive/speculator.py#L284-L295",
          "purpose": "draft prefill reuses target attention metadata and slot mappings under an identical-layout assumption"
        },
        {
          "url": "https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/attention/backends/triton_attn.py#L248-L342",
          "purpose": "Triton advertises fp8_per_token_head, multimodal-prefix support, and sm120-compatible capability support"
        },
        {
          "url": "https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/attention/backends/triton_attn.py#L439-L452",
          "purpose": "Triton requires SM89 or newer for FP8 KV; sm120 satisfies the source gate"
        },
        {
          "url": "https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/model_executor/layers/quantization/kv_cache.py#L66-L84",
          "purpose": "per-token/head KV scales are computed dynamically rather than loaded from checkpoint scalars"
        },
        {
          "url": "https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/attention/ops/triton_reshape_and_cache_flash.py#L253-L335",
          "purpose": "dynamic per-token/head FP8 quantization and scale-cache writes"
        },
        {
          "url": "https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/engine/arg_utils.py#L962-L970",
          "purpose": "v0.29 target-backend CLI accepts --attention-backend"
        },
        {
          "url": "https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/tests/quantization/test_per_token_kv_cache.py#L149-L257",
          "purpose": "generic upstream kernel coverage; not proof of the exact model, hybrid-cache, MTP3, sm120, or image combination"
        }
      ]
    },
    "observed_prior_mtp_gate": {
      "scope": "the prior 16K MTP3 candidate before the target-Triton BF16/FP8 comparison",
      "mmlu_pro_diagnostic": {
        "correct_attempts": 27,
        "attempts": 30,
        "finish_reason_counts": {
          "stop": 30
        },
        "failure": "three incorrect attempts on the computer-science item",
        "qualification_boundary": "small repeated diagnostic gate; not a full MMLU-Pro score or promotion comparison"
      },
      "vision": {
        "passed_attempts": 18,
        "attempts": 18
      },
      "interpretation": "This proves that the prior MTP3 candidate cleared its retained functional gates. It does not prove that the nested speculative attention_backend selected Triton, and it does not validate dynamic FP8 KV."
    },
    "candidate_instructions": {
      "control": {
        "config": "configs/thinkingcap-awq-vllm029-5090-16k-mtp3-triton-bf16.toml",
        "target_attention_backend": "TRITON_ATTN",
        "target_kv_cache_dtype": "bfloat16",
        "draft_kv_cache_dtype": "unset; inherit target"
      },
      "dynamic_fp8": {
        "config": "configs/thinkingcap-awq-vllm029-5090-16k-mtp3-triton-fp8.toml",
        "target_attention_backend": "TRITON_ATTN",
        "target_kv_cache_dtype": "fp8_per_token_head",
        "draft_kv_cache_dtype": "unset; inherit target",
        "scale_calculation": "automatic per token and head in the Triton cache-write kernel; no calculate-kv-scales flag"
      },
      "comparison_contract": [
        "Run the BF16 Triton control before the FP8 candidate.",
        "Keep model revision, runtime image, 16K context, C1, graph mode, MTP3, multimodal limits, prompts, and measurement controls identical.",
        "Change only the KV cache dtype between the two target-Triton arms.",
        "Leave speculative_config.kv_cache_dtype unset in both arms. An explicit BF16 draft override would create a mixed target/draft cache layout and reduce the capacity benefit.",
        "Capture resolved startup backend and cache dtype evidence. Do not attribute draft Triton selection to the nested key.",
        "Require text, reasoning, tools, vision, cache-capacity, and near-limit context evidence before retaining an FP8 recommendation."
      ]
    },
    "findings_summary": [
      {
        "severity": "high",
        "classification": "confirmed",
        "finding": "The existing nested speculative attention_backend requests TRITON_ATTN, but the exact pinned standard MTP path does not apply that field through load_eagle_model. Prior successful MTP startup and quality gates therefore do not prove a Triton drafter."
      },
      {
        "severity": "medium",
        "classification": "confirmed",
        "finding": "An explicit top-level --attention-backend TRITON_ATTN selects Triton for the target and, through the pinned standard MTP inheritance path, for the MTP head. This is the supported basis for the matched BF16 and FP8 candidates."
      },
      {
        "severity": "medium",
        "classification": "plausible",
        "finding": "TRITON_ATTN with fp8_per_token_head is a viable source-admitted candidate on sm120, but no pinned upstream integration test proves the exact Qwen3.5 hybrid, MTP3, graph, multimodal combination. Local startup, correctness, quality, and capacity evidence remain mandatory."
      }
    ],
    "breakage_probes": [
      {
        "axis": "fail-closed behavior when state or tools are unreadable",
        "input_or_state": "startup evidence lacks a model-tag or layer-specific resolved draft-backend record",
        "wrong_result": "claiming the nested key proved a Triton drafter",
        "classification": "confirmed"
      },
      {
        "axis": "malformed and boundary input",
        "input_or_state": "invalid KV dtype or FP8 KV on a pre-SM89 device",
        "wrong_result": "silent acceptance or fallback",
        "classification": "refuted",
        "evidence": "the pinned config constrains CacheDType and Triton raises below SM89"
      },
      {
        "axis": "resource exhaustion or missing bounds",
        "input_or_state": "assuming dynamic FP8 guarantees a larger qualified context from source-level byte savings alone",
        "wrong_result": "advertising a context that hybrid Mamba state, graph memory, or runtime workspace prevents",
        "classification": "plausible"
      },
      {
        "axis": "state drift across config and runtime seams",
        "input_or_state": "target remains automatically selected FlashAttention while the recipe labels only the nested MTP backend as Triton",
        "wrong_result": "recorded recipe identity differs from the pinned effective runtime path",
        "classification": "confirmed"
      }
    ],
    "missing_tests": [
      "No pinned upstream integration test covers Qwen3.5 hybrid plus MTP3 plus fp8_per_token_head plus sm120.",
      "No pinned upstream integration test covers an image request under that exact combination.",
      "The generic per-token/head kernel tests cover component behavior but cannot establish model-level correctness or quality."
    ]
  },
  "advisory_priors": [
    "The prior MTP3 candidate passed 27/30 repeated MMLU-Pro diagnostic attempts with all 30 ending in stop; the three incorrect attempts were the computer-science item.",
    "The prior MTP3 candidate passed 18/18 retained vision attempts.",
    "These observations justify a bounded target-Triton comparison but do not qualify the FP8 candidate or authorize promotion."
  ],
  "recommendation": "Run the prepared 16K target-Triton BF16 control, then the otherwise matched fp8_per_token_head candidate with draft KV inheritance. Retain FP8 only if startup identity, functional and vision gates, quality diagnostics, measured cache capacity, and bounded performance evidence pass.",
  "human_gate_required": false,
  "promoted": false
}