feat(model): Phase 1.A MTP head detection + scaffolding#171
Merged
Conversation
First step of the spec-decode MTP wiring project documented in
docs/superpowers/specs/2026-05-14-mtp-wiring-design.md.
Detects model_mtp.safetensors sidecar (DeepSeek-V3-family models — Qwen3.6,
DeepSeek V3) when loading SafeTensors models. Phase 1.A scope is detection
+ metadata only; tensor load and forward path come in Phases 1.B-5.
## What this ships
- `src/model/mtp_head.h` — new MtpHeadInfo struct (path, file size, tensor count)
- `Model::mtp_info_` — std::optional<MtpHeadInfo> field, populated when detected
- safetensors_loader.cpp — detection block after main load succeeds:
reads file size + safetensors header → tensor count, logs "MTP head detected"
- Design spec: docs/superpowers/specs/2026-05-14-mtp-wiring-design.md
(full 5-phase roadmap including the BF16 weight layout, forward kernel design,
verify-loop integration, and acceptance-rate validation plan)
## Validation
Qwen3.6-NVFP4 model load:
[INFO] MTP head detected: /models/Qwen3.6-35B-A3B-NVFP4/model_mtp.safetensors
(1.57 GiB, 19 tensors) — not yet wired (see docs/superpowers/specs/...)
Bench unchanged (pp=64 tok/s, tg=179 tok/s same as before). verify-fast green
(decode -2.28% within 3% threshold, prefill +3.26%). No production behavior
change for any model.
## What this does NOT ship
- Loading mtp.* tensors (llm_compressor_loader.cpp still skips the prefix)
- Forward pass kernel
- CLI flag / config wiring
- VRAM budget changes
- Production use
These are Phases 1.B through 5 — each scoped in the design spec, each
1-5 dev-days, total 2-3 dev-weeks for production MTP spec-decode.
## Background
Per memory `spec_decode_qwen36_broken_2026_05_02`: imp's current self-speculative
decoding has ≈0% acceptance because it reuses the trained LM head on an
intermediate hidden state with no early-exit adapter. Qwen3.6 (and DeepSeek V3)
ship a TRAINED MTP head as `model_mtp.safetensors` — this PR is the foundation
for routing that head through imp's spec-decode infrastructure.
kekzl
added a commit
that referenced
this pull request
May 14, 2026
…lding Builds on Phase 1.A (PR #171 detection) to ship a complete MTP scaffolding stack: tensors load to GPU, reduced forward kernel runs, engine API exists, CLI flag works, model produces output without crashes on Qwen3.6-NVFP4. ## Per-phase deliverables **Phase 1.B — Tensor loading** - `MtpHead` struct expanded from metadata-only to 19 named Tensor fields (pre_fc_norm_*, fc, input/post_attention_layernorm, q/k/v/o_proj, q/k_norm, router, experts_gate_up_packed, experts_down_packed, shared_expert_*, final_norm) - `safetensors_loader::load_safetensors` runs a separate load pass on `model_mtp.safetensors` after main load, dispatches the 19 tensors to MtpHead fields by name, retains the mmap via `Model::split_mmaps_` - Translation NOT applied to MTP names (literal `mtp.*` preserved) **Phase 1.C — Storage decision** - BF16 retained on disk, converted to FP16 on GPU upload (matches main weights path). NVFP4 quant deferred — 1.6 GB FP16 cost is acceptable on a 32 GB GPU running a 35 B model. **Phase 2.1 — Reduced forward kernel** (`src/runtime/mtp_forward.{cu,h}`) - `mtp_draft_step()`: emb → pre_fc_norm × 2 → concat → fc → final_norm → lm_head → argmax - Workspace alloc/free helpers (`MtpDraftWorkspace`) - **Phase 2.1 limitation**: transformer block (attention + 256-expert MoE) is SKIPPED. Compute is a passthrough of fc_out. Production correctness requires the full block (Phase 2.2 — genuinely multi-week to write the MoE forward from scratch). Acceptance rate will be far below trained optimum until Phase 2.2 lands. **Phase 3 — Engine API** - `Engine::enable_mtp_spec_decode(int k)` + `Engine::mtp_draft_one(...)` - `Engine::mtp_ws_storage_` field (type-erased to avoid header include) - Workspace allocated on enable, freed on destroy - Decode-loop auto-invocation deferred to Phase 3.5/Phase 4 production work **Phase 4 — CLI + C API** - `--mtp-spec-decode K` CLI flag - `imp_enable_mtp_spec_decode(ctx, k)` C API entry point - main.cpp calls C API after context creation if flag set **Phase 5 — End-to-end smoke test** (validated, this PR) - Qwen3.6-NVFP4 + `--mtp-spec-decode 2`: - MTP head loads (1.57 GiB, 19 tensors, BF16) - GPU upload succeeds (19 allocations) - Spec-decode enabled (k=2, hidden=2048, vocab=248320, workspace allocated) - Model produces output (125 tok/s decode, no crashes) ## Production gaps documented - Phase 2.2: full transformer block (multi-week) - Phase 3.5: decode-loop auto-invocation - Phase 5.5: acceptance rate measurement (needs 2.2 + 3.5) ## Validation - verify-fast green (decode -0.28%, prefill +1.64%) - Qwen3.6-NVFP4 smoke: end-to-end works without crashing - No production behavior change without explicit `--mtp-spec-decode` flag ## Design spec docs/superpowers/specs/2026-05-14-mtp-wiring-design.md (Phase 1.A PR #171)
github-actions Bot
pushed a commit
that referenced
this pull request
May 14, 2026
…lding (#172) Builds on Phase 1.A (PR #171 detection) to ship a complete MTP scaffolding stack: tensors load to GPU, reduced forward kernel runs, engine API exists, CLI flag works, model produces output without crashes on Qwen3.6-NVFP4. ## Per-phase deliverables **Phase 1.B — Tensor loading** - `MtpHead` struct expanded from metadata-only to 19 named Tensor fields (pre_fc_norm_*, fc, input/post_attention_layernorm, q/k/v/o_proj, q/k_norm, router, experts_gate_up_packed, experts_down_packed, shared_expert_*, final_norm) - `safetensors_loader::load_safetensors` runs a separate load pass on `model_mtp.safetensors` after main load, dispatches the 19 tensors to MtpHead fields by name, retains the mmap via `Model::split_mmaps_` - Translation NOT applied to MTP names (literal `mtp.*` preserved) **Phase 1.C — Storage decision** - BF16 retained on disk, converted to FP16 on GPU upload (matches main weights path). NVFP4 quant deferred — 1.6 GB FP16 cost is acceptable on a 32 GB GPU running a 35 B model. **Phase 2.1 — Reduced forward kernel** (`src/runtime/mtp_forward.{cu,h}`) - `mtp_draft_step()`: emb → pre_fc_norm × 2 → concat → fc → final_norm → lm_head → argmax - Workspace alloc/free helpers (`MtpDraftWorkspace`) - **Phase 2.1 limitation**: transformer block (attention + 256-expert MoE) is SKIPPED. Compute is a passthrough of fc_out. Production correctness requires the full block (Phase 2.2 — genuinely multi-week to write the MoE forward from scratch). Acceptance rate will be far below trained optimum until Phase 2.2 lands. **Phase 3 — Engine API** - `Engine::enable_mtp_spec_decode(int k)` + `Engine::mtp_draft_one(...)` - `Engine::mtp_ws_storage_` field (type-erased to avoid header include) - Workspace allocated on enable, freed on destroy - Decode-loop auto-invocation deferred to Phase 3.5/Phase 4 production work **Phase 4 — CLI + C API** - `--mtp-spec-decode K` CLI flag - `imp_enable_mtp_spec_decode(ctx, k)` C API entry point - main.cpp calls C API after context creation if flag set **Phase 5 — End-to-end smoke test** (validated, this PR) - Qwen3.6-NVFP4 + `--mtp-spec-decode 2`: - MTP head loads (1.57 GiB, 19 tensors, BF16) - GPU upload succeeds (19 allocations) - Spec-decode enabled (k=2, hidden=2048, vocab=248320, workspace allocated) - Model produces output (125 tok/s decode, no crashes) ## Production gaps documented - Phase 2.2: full transformer block (multi-week) - Phase 3.5: decode-loop auto-invocation - Phase 5.5: acceptance rate measurement (needs 2.2 + 3.5) ## Validation - verify-fast green (decode -0.28%, prefill +1.64%) - Qwen3.6-NVFP4 smoke: end-to-end works without crashing - No production behavior change without explicit `--mtp-spec-decode` flag ## Design spec docs/superpowers/specs/2026-05-14-mtp-wiring-design.md (Phase 1.A PR #171)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
First step of the spec-decode MTP wiring project. Detects `model_mtp.safetensors` sidecar (DeepSeek-V3-family models — Qwen3.6, DeepSeek V3) when loading SafeTensors models. Phase 1.A scope is detection + metadata only; tensor load and forward path come in Phases 1.B-5 (see spec doc).
What this ships
Validation
Qwen3.6-NVFP4 model load logs:
```
[INFO] MTP head detected: /models/Qwen3.6-35B-A3B-NVFP4/model_mtp.safetensors
(1.57 GiB, 19 tensors) — not yet wired (see docs/superpowers/specs/...)
```
Bench unchanged (Qwen3.6-NVFP4 pp=64=2034, tg=179 tok/s same as before). verify-fast green (decode +0.50%, prefill +1.71%).
What this does NOT ship
These are Phases 1.B-5, each 1-5 dev-days, total 2-3 dev-weeks for production MTP spec-decode (scoped in design spec).
Background
Per memory `spec_decode_qwen36_broken_2026_05_02`: imp's current self-speculative decoding has ≈0% acceptance because it reuses the trained LM head on an intermediate hidden state with no early-exit adapter. Qwen3.6 (and DeepSeek V3) ship a trained MTP head as `model_mtp.safetensors` — this PR is the foundation for routing that head through imp's spec-decode infrastructure.
Test plan
🤖 Generated with Claude Code