Downloads · 30 days
0
rwmacy/qwen3.8-27b-mtp-head-v8-b70
qwen3.8-27b-mtp-head-v8-b70 is a machine learning model from rwmacy. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for specforge. The card lists the license as apache-2.0.
Retrained native MTP (multi-token-prediction) head for Qwen3.8-27B-FP8 (hybrid 16 full-attention + 48 GDN linear-attention). Drop-in replacement for the checkpoint's shipped mtp.safetensors at /models/target/, served…
Downloads · 30 days
0
Access
Public
Updated Aug 21, 2026
Repo size
1.3 GB
Likes
1
Public
Click a slice to open those files.
.safetensors1.3 GB · 100%
From the Hugging Face model README
Retrained native MTP (multi-token-prediction) head for Qwen3.8-27B-FP8
(hybrid 16 full-attention + 48 GDN linear-attention). Drop-in replacement
for the checkpoint's shipped mtp.safetensors at /models/target/,
served with vLLM --speculative-config '{"method":"mtp","num_speculative_tokens":3}'
(k=3). Same DeepSeek-V3-style single-layer MTP structure and 22-tensor
layout as the original head; shares the target's embeddings and lm_head.
| File | Size | What |
|---|---|---|
mtp-bf16.safetensors | 849,400,392 B | BF16 master (training output) |
mtp.safetensors | 477,202,224 B | FP8 block-quantized serving drop-in — manifest block-identical to the original head |
train-v8.yaml | SpecForge training config | |
manifest.json | Training-corpus statistics |
sha256:
mtp-bf16.safetensors e1efd934f6c5b33bad5ddcc0d6cc88f1dc9c5172f375e82520d798ac9bcf8e07mtp.safetensors 19173dd6bfc310925043fcabc3dc8e41fa8a7d75c0f5e4548bc9d846d6a5cbb6crc32(id) % 20 == 0. See manifest.json.eagle3 TTT with ttt_length: 4 (draft positions
0..3, mirroring k=3 serving + bonus token), max_length: 2048,
batch 1 × accum 8, lr 5e-6 (warm fine-tune, 10× below a from-scratch run),
cosine-free, max_grad_norm 1.0, seed 42. Full config in train-v8.yaml.weight_scale_inv), so the
serving engine loads it with zero config changes.Validated as v15rc1 on the full production config (262,144-token window,
MTP k=3 + QDS draft sampling, TP=2 across two Arc Pro B70):
| Gate | Result |
|---|---|
| Train accs (final) | 0.968 / 0.931 / 0.879 / 0.859 (mean acceptance 0.889) |
| Holdout (425 never-trained prompts) | 0 errors, mean acceptance length L 2.801, 65.6 tok/s |
| Real-workload replay A/B vs the original head (v13) | parity — +0.20 tok/s mean, 14/28 wins, L 3.01 vs 3.00 (no regression, no gain) |
| Needle retrieval | 3/3 exact at 10/50/90% depth of 250k tokens |
| Boundary decode tok/s | 1k 52.6 / 8k 40.7 / 16k 33.9 / 40k 29.0 / 131k 20.8 / 250k 15.3 — every row ≥ the original head's ship table |
| 50-request soak | zero errors |
Read the replay row before deploying with expectations: on real multi-turn agent-traffic replay this head is a wash vs the original (the shipped head was already at its ceiling on that mix). The wins are in-distribution/holdout acceptance and non-regression everywhere else.
Built for and validated on ghcr.io/rmacy/qwen38-fp8-dspark:v15 — an
Intel XPU vLLM 0.21.1.dev0 fork image (self-contained: target weights,
this head, tool/reasoning parsers). Companion drafter-family repo:
rwmacy/qwen3.8-27b-dflash-drafter-fp8-b70.
Apache-2.0, matching the sibling drafter repo.