[zhuyuhua-v]

Motivation

The MiniMax-M3.1 DSpark draft builds its quantization from its own checkpoint block only (_draft_quant_config, MXFP8 [1, 32]), never from --online_quant_config. So with --online_quant_config '{"global_quant_config": "ptpc_fp8", ...}':

• the target is re-quantized to per-token FP8;
• the draft linears stay MXFP8, which is a triton block-scale GEMM (_gemm_afp8wfp8) plus a separate quantize_fp8 per linear.

Decode trace of MiniMax-M3.1-Flash TP4 on MI355X, per propose_dspark span, averaged over 400 spans:

draft linears	GEMMs per span	GEMM time per span
MXFP8 (draft ignores the flag), c16 trace	20 x _gemm_afp8wfp8	510.8 us (+ 90.3 us quantize_fp8)
per-token FP8 (draft re-quantized), c32 trace	20 x kernel_gemm_0	216.3 us (+ 85.9 us per-token quant)

The c32 trace runs larger draft batches than the c16 one, so the saving is, if anything, understated.

Change

_draft_quant_config(config, online_quant_config) now gets the server's --online_quant_config. The draft is re-quantized at load the same way the target is.

• No new flag: without --online_quant_config nothing changes.
• The draft's checkpoint validation is unchanged: MXFP8 [1, 32] or BF16.
• Test: tests/test_dspark.py::test_minimax_m31_draft_follows_online_quant_config covers both cases (with ptpc_fp8 the draft is re-quantized per token; without it, it stays MXFP8).

Validation

MiniMax-M3.1-Flash-Quark-MXFP4, TP4, MI355X, fp8 KV/index, --online_quant_config ptpc_fp8, gpu-memory-utilization 0.9, image rocm/atom-dev:nightly_202610261558. Mono is off on main with DSpark.

A/B on this branch's code: main 035a78a with and without this change, same launch, c1, 1200 s per arm, synthetic AL 4.04.

draft	output tok/s/GPU	p90 ITL	P90 interactivity	TTFT p50
MXFP8 (main)	41.5	2.60 ms	384.9	303 ms
per-token FP8 (this PR)	42.6 (+2.7%)	2.45 ms (-5.8%)	408.4 (+6.1%)	310 ms

Earlier A/B with the same re-quantization, behind an environment flag in the first revision of this PR. Main 481c2bf, c1, 1200 s per arm:

draft	output tok/s/GPU	p50 ITL	p90 ITL	P90 interactivity
MXFP8	41.0	2.31 ms	2.61 ms	383.5
per-token FP8	42.3 (+3.2%)	2.18 ms (-5.6%)	2.45 ms (-6.1%)	407.7 (+6.3%)

Accuracy. lm-eval gsm8k as InferenceX CI runs it: 5-shot, chat template, temperature 0, all 1319 samples, real acceptance, draft re-quantized.

• First revision: strict-match 0.9598 +- 0.0054 on main.
• This revision: strict-match 0.9636 +- 0.0052.
• SGLang CI on MI355X: 0.9598 +- 0.0054.

🤖 Generated with Claude Code

[github-actions]

🏷️ CI Guide

Runs automatically on every eligible PR before approval:

• ✅ Pre Checkin: Black, Ruff, catalog schema validation, non-GPU unit tests

Heavy model tests:

• ✅ Run after the PR is approved and Pre Checkin passes
• ✅ Run immediately when an approval review is submitted
• ✅ Can be requested before approval with labels

Label	Tests
ci:full	Run all heavy PR model tests: native ATOM, vLLM, and SGLang
ci:atom	Run native ATOM model accuracy tests
ci:vllm	Run ATOM vLLM OOT model accuracy tests
ci:sglang	Run ATOM SGLang model accuracy tests

Heavy jobs are skipped when the PR is not approved and no matching ci:* label is present.
Add labels via the sidebar or gh pr edit 2743 --add-label <label>

[Copilot]

🔵 Needs a closer look

The inference-numerics and GPU GEMM dispatch changes require validation beyond source-only review, with current-revision GSM8K accuracy and GPU A/B results still pending.

No open issues 🎉

What changed in this PR

This PR makes MiniMax-M3.1’s DSpark draft honor ATOM’s existing --online_quant_config, aligning its load-time re-quantization with the target model.

Changes:

• Pass the server’s online quantization configuration into draft quantization setup.
• Preserve checkpoint validation and existing behavior when the option is absent.
• Add a parameterized regression test covering default MXFP8 and ptpc_fp8 configurations.

File	Description
tests/​test_dspark.py	Tests draft quantization with and without the online override.
atom/​models/​minimax_m3/​dspark.py	Forwards online quantization settings to the draft’s quantization configuration.

_{Review effort: Balanced}

[zhuyuhua-v]

drop this pr since SA won't allow re-quant in draft models.