An EAGLE-3 draft (speculator) for accelerating Korean generation from gemma-4-26B-A4B-it (Gemma 4 MoE; 26B total, ~4B active) via speculative decoding. The public RedHat draft is English-only (Magpie + UltraChat) and accepts Korean tokens poorly, so this draft was retrained on Korean prompts with on-policy responses regenerated by the verifier.

Method: EAGLE-3 (vllm-project/speculators)
Verifier (target): RedHatAI/gemma-4-26B-A4B-it-FP8-Dynamic (FP8 MoE; serving + hidden-state extraction)
Warm start: RedHatAI/gemma-4-26B-A4B-it-speculator.eagle3
Training data: ~150k prompts sampled from sh2orc/bccard-maywell-jojo0217-markai-lcw99-kendamarron-microsoft (1.71M-row Korean QA; instruction column only). Answers regenerated on-policy by the verifier (text-only).

Serving (vLLM)

bash

VLLM_USE_FLASHINFER_SAMPLER=0 vllm serve RedHatAI/gemma-4-26B-A4B-it-FP8-Dynamic -tp 1 \
  --max-model-len 4096 \
  --limit-mm-per-prompt '{"image":0,"audio":0,"video":0}' \
  --speculative-config '{
    "model": "<this-repo-id>",
    "num_speculative_tokens": 3,
    "method": "eagle3",
    "draft_tensor_parallel_size": 1
  }'

Tune num_speculative_tokens (3–6) on measured acceptance / TPS. The draft uses the verifier's tokenizer and chat template. The 26B-A4B is multimodal; for text workloads pass --limit-mm-per-prompt to disable image/audio/video.

Notes for the MoE target

The verifier is a Mixture-of-Experts model (128 experts, top-8). On some Blackwell parts (sm_120 / sm_121) the FP8 fused-MoE Triton kernel can exceed shared memory; if serving fails with an out of resource: shared memory error, use a vLLM MoE fallback config or the NVFP4 path.
Acceptance is measured against RedHatAI/gemma-4-26B-A4B-it-FP8-Dynamic. Pairing with a different target (or a base, non--it model) will change results.

Limitations

Trained on general Korean QA. Domain-specific traffic (e.g. finance) may benefit from another training cycle on domain-matched data.

License

Apache 2.0. The base Gemma 4 (Apache 2.0), the RedHat FP8 verifier (Apache 2.0), and the RedHat EAGLE-3 warm-start checkpoint (Apache 2.0) are all Apache 2.0, so this draft is released under Apache 2.0 as well. (Informational, not legal advice.)

BCCard/MoAI-gemma-4-26B-A4B-it-speculator.eagle3

gemma-4-26B-A4B-it EAGLE-3 Draft (Korean-optimized)

Serving (vLLM)

Notes for the MoE target

Limitations

License

BCCard/
MoAI-gemma-4-26B-A4B-it-speculator.eagle3