Datasets:
First EAGLE-3 draft head for Hermes-4.3-36B, epoch 1: converged at 39% of the epoch, lands at 1.29–1.57x against a 1.7–2.2x precedent band
Rig: one RTX 5090 32GB · sglang 0.5.14 · teacher cyankiwi/Hermes-4.3-36B-AWQ-4bit (Seed-OSS) · greedy, batch 1 Status: epoch 1 complete (2026-07-17). Weights + model card follow with the release; this is the rig report.
What this is
Hermes-4.3-36B (Seed-OSS architecture) had no speculative-decoding draft head on all of Hugging Face — the architecture isn't wired for EAGLE-3 in any serving engine, and the standard SpecForge recipe assumes multi-GPU offline training with terabytes of cached teacher activations. This run trained the first one, entirely on one consumer card, using SpecForge's online mode (teacher resident in VRAM, activations recomputed per batch — zero disk cost) with a 4-bit AWQ teacher.
Making it fit took six patches on top of SpecForge (branch: single-GPU online):
- Optimizer CPU-offload — fp32 masters + AdamW moments (8.4 GiB) live in host RAM, step on CPU (~0.5 s, fully amortized under a ~3.7 s teacher-bound step)
- Draft model built on CPU after the engine claims its VRAM floor
- Skip FSDP at world_size 1 — FSDP silently allocates a 1.48 GiB gradient for the frozen embedding
- Frozen embedding in host RAM (1.6 GiB; one lookup per batch)
- Chunked acceptance-metric softmax (memory flat in sequence length)
- Gradient checkpointing over the TTT unroll via flex_attention — unlocked 2048-token samples from night 2
Five of these are filed upstream (SpecForge #669, #670, #671), plus the Seed-OSS serving port to sglang (#30930) and EAGLE-3 hooks for vLLM (#48403).
Training: 54K curated conversations (34K Hermes-3 + 20K UltraChat, tool-calls excluded), one epoch = 54,000 steps across 8 unattended nights (~55 GPU-hours, every night rc=0). Night 1 at max-length 1024; nights 2–8 at 2048 after the gradient-checkpointing patch. Pace ~3.7 s/step at both lengths — the step is teacher-prefill-bound, so doubling the sample length was free.
The release bench
Fixed prompt set (v1, 8 prompts per workload, never edited in place), 256-token completions, greedy,
cuda graphs ON, mem_frac 0.85, ctx 4096. Baseline is the same engine, same checkpoint, no speculative
decoding: 66.2–66.3 tok/s on every workload.
| config | prose | code | repetitive | chat |
|---|---|---|---|---|
| chain-3-1-4 | 1.22x (1.54) | 1.20x (1.52) | 1.38x (1.75) | 1.13x (1.42) |
| tree-3-4-8 | 1.38x (1.81) | 1.38x (1.80) | 1.57x (2.06) | 1.29x (1.69) |
| tree-5-8-16 | 1.32x (1.96) | 1.33x (1.98) | 1.48x (2.22) | 1.23x (1.84) |
(accept-len in parentheses)
tree-3-4-8 wins all four workloads and is the release config. The config worth studying is tree-5-8-16: it accepts deeper than 3-4-8 on every workload (up to 2.22 on repetitive) and still delivers less speedup, because the wider tree's draft cost eats more than the extra accepted tokens return. Acceptance is not speedup: every drafted token costs teacher-side compute whether it lands or not.
Finding 1: the head converged at 39% of the epoch
The release bench was dry-run at step 21,000, where tree-3-4-8 already won every workload at 1.30–1.53x. At epoch end, 33,000 steps later, the same config lands at 1.29–1.57x — the final checkpoint reproduces the 39%-of-epoch checkpoint within noise.
The nightly fixed smoke (chain 3/1/4, n=1 — a trend instrument, not a release number) says the same thing with eight points: accept-lens froze at night 3 (code 1.67, repetitive 1.46, prose 1.38) and every subsequent point stayed within smoke noise. The last five nights, 60% of the epoch, moved nothing. The data is saturated for this head.
Finding 2: the precedent band stays out of reach, and the mechanism is visible
Trained-head precedent for EAGLE-3 heads is 1.7–2.2x. This head lands at 1.29–1.57x. The gap is not undertraining (finding 1 rules that out on this data). The arithmetic: at accept-len 1.80, a free drafter would deliver ~1.8x; the measured 1.38x means roughly a quarter of the theoretical gain goes to running the draft head itself, at batch 1 on one card where the verify pass has no batching to amortise against. Deeper trees make this worse, not better (the tree-5-8-16 row above). The levers that could close the gap are a cheaper draft pass or training data that pushes acceptance past ~2.5 — not more epochs of this data.
Decision
No epoch 2 on the same data — the curve is flat and re-training on saturated data spends GPU-nights on nothing. The head releases as-is with honest numbers: 1.3–1.6x real decode speedup on a single RTX 5090, tree-3-4-8, batch 1. A v2, if it happens, changes the data mix, not the step count.
Honest caveats
- 8 prompts per workload, greedy, 256-token completions, single run — the fixed set makes points comparable across the run, not statistically tight. Curve shape (8 nightly points) is the evidence; any single cell has smoke-level noise.
- On the repetitive workload, 6 of 8 baseline completions terminate naturally before 256 tokens — the workload's tok/s rests on fewer decoded tokens than the others.
- One teacher quant (AWQ 4-bit), one engine (sglang 0.5.14), batch 1 only. Server-style batched serving changes speculative-decoding economics entirely; these numbers are the local-single-user story.
- The 1.7–2.2x band is precedent from published trained heads (different teachers, data recipes, and hardware) — it's the reference the run was read against, not a controlled comparison.