README.md

---
base_model: 0xSero/DeepSeek-V4-Flash-0731-REAP
tags:
- speculative-decoding
- draft-model
- mtp
- jetson
library_name: safetensors
---

# dspark-mtp-draft-head-0731reap-v1 — DSpark MTP draft head for `0xSero/DeepSeek-V4-Flash-0731-REAP`

This is **not a standalone language model.** It is a fine-tuned set of **DSpark MTP draft-head
tensors** for `0xSero/DeepSeek-V4-Flash-0731-REAP`, used as the speculator in self-speculative decoding. It only does
anything in combination with the base checkpoint.

## What it is for

Speculative decoding with **greedy verification**: the draft head proposes several tokens, the base
model verifies them in one pass, and the longest correct prefix is accepted. Because verification is
greedy and exact, **the emitted token sequence is identical to what the base model would have
produced on its own** — this is a latency optimisation, not a quality trade. Every head in this line
is gated on that property before it is measured (`LOSSLESS GATE: first 8 tokens match base AR`), and
a head that changes the output is rejected regardless of how fast it is.

## Measured

Hardware: Jetson AGX Thor (`sm_110a`, 20 SMs, 122.8 GiB unified LPDDR5X, CUDA 13.0), custom CUDA
inference engine, block size 6.

| | tok/s |
|---|---|
| base autoregressive decode (no speculation) | 13.8 |
| stock DSpark head shipped with the checkpoint | 22.66 |
| **this head** | **25.5312** |

Acceptance tau (tokens per verify, 8-prompt suite mean): **3.8438** out of a maximum of 7.

**Protocol, which is part of the number.** 8-prompt suite spanning agentic_format, code_edit,
code_gen, explanation, long_context, multi_turn, reasoning and short_factual; NGEN0 >= 200 generated
tokens (acceptance is a transient below ~128 tokens and a short-generation figure is not comparable
to anything); clocks pinned; page cache dropped; no profiling instruments in the binary; adaptK
1.50. Measurements on this box have a residual sd of ~0.5 % per run with a discarded warm-up and
shuffled run order — differences below ~1 % are not meaningful without replicates.

## Limitations, measured rather than assumed

1. **It regresses on held-out continuation drafting.** Against a true paired control — the stock head
   over the same prompts, budget and threshold — the fine-tuned heads are *slower* at drafting the
   model's own long continuations (-0.40 tau for the first session's head, -0.65 for the second).
   They win the mixed suite and lose this. Both numbers are real; which one matters depends on the
   workload.
2. **Training helps weak categories and hurts strong ones.** Rank-ordered across the suite: the
   categories the stock head was worst at improved most, and the ones it was best at regressed.
   Balancing the training corpus repaired some of that (`short_factual`, half of `code_edit`) and
   left `long_context` and `agentic_format` untouched.
3. **Tuned for one engine and one checkpoint.** The draft head taps specific backbone layers and is
   fine-tuned against activations captured from this exact checkpoint. It is not expected to
   transfer to another quantisation, another REAP revision, or another serving stack.

## Files

- `config.json` (0 MB)
- `eval.log` (0 MB) — the raw measurement log behind the numbers above
- `head_card.json` (0 MB) — the archive record written when this head was measured
- `mtp_trained.safetensors` (1035 MB) — the trained MTP draft-head tensors — the artifact
- `train_metrics.json` (0 MB) — per-step training loss history

## How to use it

The loadable head is the base checkpoint with these tensors substituted in, re-quantised to the
formats the checkpoint uses (fp8 e4m3 with F8_E8M0 128x128 block scales for the projections, bf16
and fp32 for the rest). Materialise it locally:

```bash
git clone https://github.com/<your-org>/deepseek-v4-flash-0731-cuda
python3 tools/build_trained_head.py \
    --base /path/to/DeepSeek-V4-Flash-0731-REAP \
    --trained mtp_trained.safetensors \
    --out /path/to/head_out
```

The build is deterministic and self-checked: it recomputes each block scale from the trained values
(reusing the old scale against new weights is the silent way to get a plausible tensor with the
wrong magnitude), snaps that scale to an exact power of two *before* quantising because E8M0 stores
a bare exponent, and refuses to write any tensor whose round-trip error exceeds 0.10. Worst observed
round-trip error for this head: see `provenance.json`.

## Provenance

Base model `0xSero/DeepSeek-V4-Flash-0731-REAP`, used as shipped — no re-pruning and no additional quantisation. Trained
on activations captured from that checkpoint over the session's prompt continuations the base model itself
generated. Full training and measurement history, including the rejected candidates and the
corrections, is in `LOOP_LOG.md` and `S5_PROGRESSION.md` in the engine repository.

Engine revision at measurement time: `85dbea6cf`.