[None][test] Switch MiniMax-M3 Eagle3 accuracy test to the GQA draft head

[zheyuf]

Description

Switches TestMiniMaxM3::test_nvfp4_eagle3 from the MHA EAGLE3 draft head
(Inferact/MiniMax-M3-EAGLE3) to the GQA retrain
(Inferact/MiniMax-M3-EAGLE3-GQA).

Per the model card the GQA head is trained on the same data (kimi-mtp, OpenCodeInstruct, SWE-bench,
SWE-bench-Pro); the only architectural change is the draft attention, num_key_value_heads 64 → 4.

Why

The EAGLE3 draft is a single dense attention layer, so its KV cache is disproportionately expensive
next to MiniMax-M3's sparse-attention target. From the published configs (head_dim 128, BF16 KV),
draft KV per token per rank at TP4:

	draft KV per token/rank
MHA (num_key_value_heads 64)	64/4 × 128 × 2 × 2 B = 8192 B
GQA (num_key_value_heads 4)	4/4 × 128 × 2 × 2 B = 512 B

A 16× smaller draft cache leaves substantially more of the KV pool for the target, and in
disaggregated serving it proportionally shrinks the draft share of the KV transfer.

Acceptance is comparable: the GQA card reports MT-Bench 2.668 / 55.62% and SPEED-Bench (qualitative)
2.561 / 52.04% mean accepted length / accept rate at num_speculative_tokens=3.

Why no runtime change is needed

The draft KV cache manager and the drafter's attention both read num_key_value_heads from the
drafter's own config, not the target's:

• SpecDecOneEngineForCausalLM builds the draft config via
  ModelConfig.from_pretrained(spec_config.speculative_model) (modeling_speculative.py)
• Eagle3Attention passes config.num_key_value_heads straight into Attention
  (modeling_speculative.py)
• KvCacheCreator._get_effective_draft_config() returns that draft config, and the V2 KV cache
  manager derives per-rank KV heads from it (_util.py, kv_cache_manager_v2.py)

This is the same path the existing GQA Llama EAGLE3 heads already exercise (e.g.
EAGLE3-LLaMA3.1-Instruct-8B is 32 Q / 8 KV heads) — the MiniMax-M3 MHA head was the unusual case,
not this one.

Test Coverage

accuracy/test_llm_api_pytorch.py::TestMiniMaxM3::test_nvfp4_eagle3 — both l0_dgx_b200_m3 variants
(attention_dp False/True). Test ids are unchanged, so no test-list edits are required.

Verification done

• Swept the branch (tests/, docs/, examples/, jenkins/): this test is the only reference to
  the MHA head, on both the original and current feat/m3_with_msa.
• The MiniMax-M3 deployment guide does not pin an EAGLE3 head, so no doc change is needed.
• Confirmed the two checkpoints' configs are identical apart from num_key_value_heads, and that the
  GQA weights match it (k_proj is [512, 12288] = 4 × 128, all BF16).
• Smoke-tested a 4-GPU TP4 deployment: the GQA head loads and completes engine warmup with no code
  changes and no errors.

Acceptance thresholds

The GQA card publishes MT-Bench / SPEED-Bench but no chat-GSM8K figures, so the MHA-derived
references (rate 0.839, length 3.518) and thresholds (rate > 0.78, length > 3.3) are carried over and
labelled provisional in the test comments. They should be recalibrated from the first green CI run on
this head.

Dependencies

• MiniMax-M3-EAGLE3-GQA added to the CI model store (ftp/llm-models!583, merged) — allow the
  usual sync delay to the CI scratch before running the test