Gemma 4 E4B · vLLM · FP8 + EAGLE3

status: blocked

Configuration

Modelgoogle/gemma-4-E4B-it
CompanyGoogle
FamilyGemma
Parameters~4B effective (elastic / MatFormer)
EnginevLLM + EAGLE3 (RedHatAI speculator convention) (speculative decoding)
Quant / precisionFP8
Why this quantWould stack leon-se's FP8-Dynamic base (the done FP8 base/MTP quant) with a RedHatAI EAGLE3 speculator, mirroring the 26B-A4B/31B EAGLE3 rows — but no E4B EAGLE3 head exists to draft with (see below).
Downloadleon-se/gemma-4-E4B-it-FP8-Dynamic
Context window65536
Input modalitiestext, image (served text-only here)

Measured results

Prefill tok/s
Decode tok/s
Peak memory (GB)
Completed2026-07-02

Full run command

# BLOCKED on draft sourcing, NOT the harness — no EAGLE3 head exists for Gemma-4 E4B.
# RedHatAI publishes gemma-4 EAGLE3 speculators ONLY for 26B-A4B and 31B
# (RedHatAI/gemma-4-{26B-A4B,31B}-it-speculator.eagle3, both 200); the E4B name
# RedHatAI/gemma-4-E4B-it-speculator.eagle3 is 404, and an HF search across ALL
# authors (2026-07-02) returned NO E4B eagle3 head from anyone. So there is nothing
# to pass as the EAGLE3 draft model. Intended command if a head existed:
scripts/bench-vllm-serving.sh leon-se/gemma-4-E4B-it-FP8-Dynamic 65536 32 1000 900 256 \
  --speculative-config '{"method":"eagle3","model":"","num_speculative_tokens":3}'

gemma-4-e4b Google Gemma FP8 ≤4B conc-32

Notes

BLOCKED — no EAGLE3 head exists for Gemma-4 E4B (draft-sourcing block, not a harness/image wall)

The E4B FP8 row already has base (869.7) and +MTP (1261.5); only the EAGLE3 cell was open. Filling it needs an EAGLE3 speculator trained against the Gemma-4 E4B target — and none exists.

  • RedHatAI ships gemma-4 EAGLE3 heads for only two sizes: RedHatAI/gemma-4-26B-A4B-it-speculator.eagle3 and RedHatAI/gemma-4-31B-it-speculator.eagle3 (both verified HTTP 200, both already benchmarked here). The inferred E4B name RedHatAI/gemma-4-E4B-it-speculator.eagle3 returns 404 (as does the 12B).
  • No E4B EAGLE3 head from any other author. An HF model-search on 2026-07-02 (gemma-4-E4B eagle3 / gemma-4-E4B speculator) returned zero results — unlike the 12B, which at least has third-party heads (deepseek-ai/eagle3_gemma4_12b_ttt7, BCCard/...).
  • So this is blocked on sourcing/training a draft, not on the harness or an image wall — the FP8 target itself serves fine (see the base/MTP rows). Revisit if RedHatAI (or a trusted lab) publishes an E4B EAGLE3 speculator. The MTP path already covers E4B spec-decode on vLLM (1261.5 tok/s).