docs/qwen38-tensorfold-gap.md

Qwen3.8 Flash Next: TensorFold gap

Result#

The 2026-09-29 DGX reproduction attempt is BLOCKED_MISSING_ARTIFACTS. It publishes no performance number and no cross-engine ratio.

A fresh successful audit lease then executed the committed bounded scan script on dgx:gpu0 (NVIDIA GB10, driver 580.173.02). The raw job ID is not retained; the evidence stores only a SHA-256 lease fingerprint. The raw receipt establishes absence only from the two documented staging paths and from /workspace to depth 3 for the committed TensorFold, MiaAI, Vontra, MLX-MTP, and Flash-Next-MTP patterns. It does not claim exhaustive host-wide absence. The required artifacts/source were not found in that retained scope, so runner artifact manifests and source revisions could not be populated and measurement did not start.

Consequence#

Artifact hashes and launch environment files could not be created from measured bytes. Therefore no server was started and the executable correctness gate did not run. Serial decode, drafted decode, prefill and serving ladders, busy clock windows, memory capture, and same-tool profiles are all NOT_RUN_PREREQUISITE. Production vLLM remains the required named denominator and is recorded as NOT_RUN: this blocked discovery did not stage or inspect a runnable production-vLLM denominator.

TensorFold’s publisher rates remain unverified publisher claims. They are not local observations, a target derived from a local profile, or evidence for an optimization. No ratio, speedup, winner, residual gap, or ceiling is stated.

MTP artifact verdict#

The vllm.cpp GGUF verdict is independently BLOCKED_NO_MTP_WEIGHTS. The committed real-header manifest tests/vllm/models/qwen4_exp_gguf_manifest.inc pins revision 8bdc666649440e9bdc97e16f3f75782c98478ff5 and contains 1,224 tensor entries. It has trunk blocks 0 through 47 and zero tensor names matching mtp, nextn, draft, eh_proj, enorm, or hnorm. This verdict applies to that GGUF; it does not claim anything about the absent TensorFold MLX checkpoint.

Consequently, native MTP cannot be implemented against the selected GGUF bytes. Conversion is not authorized or justified here, and no speculative loader is scoped for weights that are absent.

Follow-on disposition#

The execution-plan outcomes are explicit:

  • Task 5 / W2: NO_PORT_DECISION. Incremental QSA is not implemented because Task 4 produced no profile proving repeated compressor work material.
  • Task 6 / W3: BLOCKED_NO_MTP_WEIGHTS. The selected GGUF has no MTP head.
  • W4-W5: SKIPPED_NO_PROFILE. Long-context selection, PLE, and prefill geometry have no measured bottleneck or correctness baseline.
  • Task 7 / W6: SYNTHESIS_COMPLETED_NO_PRODUCT_OPTIMIZATION. This record is the completed synthesis; no profile-justified product change was retained.

The campaign can resume measurement only after the runner prerequisites are available in the documented/configured staging scope. It must then rerun correctness before timing rather than reusing this blocker as benchmark evidence.

Measurement tools#

The endpoint harness guide includes a request example, required tokenizer identity, and the evidence contract. The harness measures an already-running OpenAI-compatible endpoint. For the vllm-cpp adapter, --draft off records the declared policy but does not reconfigure speculative decoding on the running server.

Current TensorFold/vllm.cpp comparisons report PROFILE_COMPARISON: they prove neither token parity nor a cross-engine ratio. Matched prompt-token fingerprints are separate from reply-text token evidence. TensorFold’s generated token IDs cannot be compared with vllm.cpp’s retokenized reply text. Check the output JSON’s refusal_reason and sample evidence. A zero exit status alone does not prove a valid measurement.

For serial capture, start with the TensorFold environment template and vllm.cpp environment template . They list the source revisions, artifact hashes, launch commands, and server lifecycle commands required by the capture runner . Fill copies outside the repository with measured local values. The runner requires an active dgx:gpu0 lease. These templates do not resolve the missing artifacts or establish a successful capture.

Evidence and validation#

Versioned evidence: .agents/evidence/bench-qwen38-tensorfold-gap/20260929T180547Z/.

Validate it with:

python3 tools/bench/validate_qwen38_tensorfold_evidence.py \
  .agents/evidence/bench-qwen38-tensorfold-gap/latest