Repository navigation
[3/4] Add MiniMax-M3.1-Flash recipes - #4583
Conversation
MiniMax-M3.1-Flash shares MiniMax-M3's architecture apart from MSA attention on every layer and a RoPE base of 1e7, so it reuses the M3 plugin and bridge. The release ships MXFP8 linears and NVFP4 routed experts; prepare dequantizes it to BF16 with tools/convert_minimax_m3_to_bf16.py before the torch_dist conversion. Adds the full-model and 4-layer recipes and their model args.
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Claude Code Review
This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.
Tip: disable this comment in your organization's Code Review settings.
Same as the MiniMax-M3 4-layer test, from CharyZeng/MiniMax-M3.1-Flash-4layer-bf16: layers 0,1,3,4 of the release dequantized to BF16 by tools/convert_minimax_m3_to_bf16.py.
|
/rerun-test tests/e2e/megatron/model_scripts/test_minimax_m3_1_flash_4layer_ci.py |
|
❌ The execution job failed; inspect the workflow run for the failing step. |
|
/rerun-test tests/e2e/megatron/model_scripts/test_minimax_m3_1_flash_4layer_ci.py |
|
✅ |
|
/rerun-test tests/e2e/megatron/model_scripts/test_minimax_m3_1_flash_4layer_ci.py |
|
❌ The execution job failed; inspect the workflow run for the failing step. |
|
/rerun-test tests/e2e/megatron/model_scripts/test_minimax_m3_1_flash_4layer_ci.py |
|
✅ |
MiniMax-M3.1-Flash shares MiniMax-M3's architecture apart from MSA attention on every layer and a RoPE base of 1e7, so it reuses the M3 plugin and bridge. The release ships MXFP8 linears and NVFP4 routed experts;
preparedequantizes it to BF16 withtools/convert_minimax_m3_to_bf16.pybefore the torch_dist conversion. Adds theMiniMax-M3.1-Flash-bf16andMiniMax-M3.1-Flash-4layer-bf16recipes and their model args.Stack: #4581 → #4582 → #4583 (this) → #4584.
Validation
MiniMax-M3.1-Flash-4layer-bf16(GB300, sglang-miles before the v0.5.23 rebase plus the weight-update fix): abs diff 0.0052 over 3 steps, weight check passes.sglang-miles(which now carries upstream M3.1-Flash support, [Feat] minimax: Support MiniMax-M3.1-Flash sgl-project/sglang#45925); will re-run the 4-layer recipe once that image includes [MiniMax-M3] Keep the fused q/k/v + index GEMM writable by online weight updates sgl-project/sglang#47612.MiniMax-M3recipe.tests/e2e/megatron/model_scripts/test_minimax_m3_1_flash_4layer_ci.pypasses on 4 × H200 (stage-c-4-gpu-h200) on the SGLang v0.5.23radixark/miles:devimage with [MiniMax-M3] Keep the fused q/k/v + index GEMM writable by online weight updates sgl-project/sglang#47612: https://github.com/radixark/miles/actions/runs/41275851020.ci-sglang-pr: #47612