Skip to content

[None][test] Switch the MiniMax-M3 Eagle3 CI to the GQA draft head #17514

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
zheyuf wants to merge 1 commit into from
+28 −15
Closed

[None][test] Switch the MiniMax-M3 Eagle3 CI to the GQA draft head #17514

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
43 changes: 28 additions & 15 deletions tests/integration/defs/accuracy/test_llm_api_pytorch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7876,7 +7876,7 @@ def _run_nvfp4_eagle3_disagg(self, model_name, model_path, max_draft_len,
speculative_config = {
"decoding_type": "Eagle3",
"max_draft_len": max_draft_len,
"speculative_model": f"{llm_models_root()}/MiniMax-M3-EAGLE3",
"speculative_model": f"{llm_models_root()}/MiniMax-M3-EAGLE3-GQA",
"eagle3_one_model": True,
}
common_config = {
Expand Down Expand Up @@ -7981,8 +7981,15 @@ def _run_nvfp4_eagle3_disagg(self, model_name, model_path, max_draft_len,

# Chat-format acceptance probe — the same workload and
# thresholds as the aggregated arm's (200 GSM8K questions,
# chat template, greedy, 512 tokens; drafter card reference
# rate 0.839 / length 3.518). The disagg fixture has no
# chat template, greedy, 512 tokens; MHA card reference rate
# 0.839 / length 3.518). Calibration for the GQA head: this
# arm measures 3.369-3.395 / 0.790-0.798 across runs while
# the aggregated arm measures 3.515 / 0.838, so the floors
# sit below the disaggregated (lower) pair — length 3.3 stays
# tight enough to catch a drafter-KV transfer regression
# (which cost the MHA head 3.44 -> 3.33 on this workload)
# while the rate floor drops to 0.76 for headroom.
# The disagg fixture has no
# get_stats, so the probe window comes from the generation
# worker's /metrics buffer (enable_iter_perf_stats), drained
# right before the probe; the dataset loads from the models
Expand Down Expand Up @@ -8026,9 +8033,9 @@ def _run_nvfp4_eagle3_disagg(self, model_name, model_path, max_draft_len,
print(f"MiniMax-M3 Eagle3 disagg chat-GSM8K acceptance: rate="
f"{chat_rate:.3f}, mean acceptance length="
f"{chat_length:.3f} ({steps} spec iterations)")
assert chat_rate > 0.78, \
assert chat_rate > 0.76, \
f"Eagle3 chat-GSM8K acceptance rate too low: " \
f"{chat_rate:.3f} (threshold 0.78, reference 0.839 from " \
f"{chat_rate:.3f} (threshold 0.76, reference 0.839 from " \
f"the drafter card)"
assert chat_length > 3.3, \
f"Eagle3 chat-GSM8K acceptance length too low: " \
Expand Down Expand Up @@ -8064,7 +8071,7 @@ def test_nvfp4_eagle3(self, tp_size, ep_size, attention_dp,
return
spec_config = Eagle3DecodingConfig(
max_draft_len=max_draft_len,
speculative_model=f"{llm_models_root()}/MiniMax-M3-EAGLE3",
speculative_model=f"{llm_models_root()}/MiniMax-M3-EAGLE3-GQA",
)
# The runtime forces tokens_per_block per implementation (128 MSA / 32
# reference).
Expand Down Expand Up @@ -8127,9 +8134,9 @@ def drain_spec_stats(llm):
task = GSM8K(model_name)
task.evaluate(llm)

# Chat-format acceptance — the drafter's training distribution
# (Inferact/MiniMax-M3-EAGLE3 card: 0.839 / 3.518). Reuses the
# live engine and the cached dataset; ~20 s under CUDA graphs.
# Chat-format acceptance — the drafter's training distribution.
# Reuses the live engine and the cached dataset; ~20 s under
# CUDA graphs.
questions = [
r["question"]
for r in load_dataset("gsm8k", "main", split="test")
Expand All @@ -8150,17 +8157,23 @@ def drain_spec_stats(llm):
assert steps > 0, "no speculative iterations recorded"
chat_rate = accepted / drafted
chat_length = 1 + accepted / steps
# Published reference (Inferact/MiniMax-M3-EAGLE3 drafter card):
# rate 0.839, length 3.518. Our testing thresholds: rate > 0.78,
# length > 3.3.
# The GQA drafter card publishes no GSM8K figure, so the
# reference stays the MHA card's (rate 0.839, length 3.518):
# the GQA head is retrained on the same data with only the
# attention changed (64 -> 4 KV heads), the cards agree on the
# benchmark they share (MT-Bench 2.698 vs 2.668), and this arm
# measures the GQA head at 0.838 / 3.515 — indistinguishable
# from the MHA reference. Floors: rate > 0.76, length > 3.3
# (see the disaggregated arm for how they are calibrated).
ref_rate, ref_length = 0.839, 3.518
ref_source = "the Inferact/MiniMax-M3-EAGLE3 drafter model card"
ref_source = ("the Inferact/MiniMax-M3-EAGLE3 (MHA) drafter "
"card, which the GQA head matches on GSM8K")
print(f"MiniMax-M3 Eagle3 chat-GSM8K acceptance: rate="
f"{chat_rate:.3f}, mean acceptance length="
f"{chat_length:.3f} ({steps} spec iterations)")
assert chat_rate > 0.78, \
assert chat_rate > 0.76, \
f"Eagle3 chat-GSM8K acceptance rate too low: {chat_rate:.3f} " \
f"(threshold 0.78, reference {ref_rate} from {ref_source})"
f"(threshold 0.76, reference {ref_rate} from {ref_source})"
assert chat_length > 3.3, \
f"Eagle3 chat-GSM8K acceptance length too low: " \
f"{chat_length:.3f} (threshold 3.3, reference {ref_length} " \
Expand Down
Loading