feat(minimax): support MiniMax-Text-01 and MiniMax-M2

[y258dd]

Summary

• Add MiniMaxForCausalLM support under csrc/models/minimax/.
• Support hybrid full-attention and lightning linear-attention layers.
• Support MiniMax-style indexed recurrent state pools for prefill and decode.
• Add the MiniMax block-sparse MoE path with a dense fallback.
• Add the _remap_minimax HuggingFace weight remapper in python/infinilm/modeling_utils.py.
• Add MiniMax operator, smoke, and HuggingFace-aligned end-to-end tests under test/models/minimax/.

Motivation

MiniMax-Text-01 and MiniMax-M2 use hybrid attention and a recurrent linear-attention state. The linear-attention path evaluates:

ratio   = exp(-slope)
state   = ratio * state + outer(k_t, v_t)
out_t   = q_t @ state

The recurrent state is read from initial_state[initial_state_indices[b]] and written to initial_state[final_state_indices[b]]. This model adaptation is needed to run MiniMax models in InfiniLM.

Closes # (N/A: internal model-support work; no tracking issue in this fork)

Type of Change

• feat — new feature / new model
• fix — bug fix
• perf — performance improvement (no behavioral change)
• refactor — code restructuring without behavior change
• test — adding or fixing tests only
• docs — documentation only
• build / ci — build system or CI configuration
• chore — tooling, formatting, or other non-code changes
• Breaking change

Test Results of Involved Models on Supported Platforms (Please attach screenshots)

Platform	Test	Result	Notes
CPU	MiniMax lightning attention operator	Passed	Matches NumPy reference
CPU	MiniMax smoke prefill/decode	Passed	State continuity error 1.16e-10
CPU	MiniMax vs HF transformers	Passed	Prefill 5.59e-4, decode 1.59e-4
NVIDIA	MiniMax model-level tests	Pending	Planned on RTX 4090 / sm_89
Ascend	New-architecture InfiniOps operator	24/24 passed	Separate InfiniOps PR #995; not this old-architecture branch

Not yet run:

• examples/test_infer.py: pending NVIDIA/real-model validation.
• examples/bench.py: no performance claim.
• test/bench/test_benchmark.py: pending.
• python/infinilm/server/inference_server.py + scripts/test_perf.py: pending.

Benchmark / Performance Impact

N/A. This PR adds model support and does not claim performance improvements.

Notes for Reviewers

• The PR targets the InfiniLM-v0.2.9c branch as requested by the administrator.
• The model supports both minimax and minimax_m2 model types.
• The linear-attention implementation requires the companion lightning_attention InfiniCore operator.
• The current CPU tests use num_experts=1, which exercises the dense fallback. The full multi-expert MoE comparison is the next validation step.
• The source state row is never modified; the destination row is written only after all tokens have been processed.
• The old-architecture NVIDIA model-level tests are still pending; this PR is intentionally opened as a draft until that validation is complete.

CI / ChatOps

CI does not run automatically on pull requests. Trigger it manually from the Actions tab, or ask a maintainer to comment /retest or /test on this PR.

---

Checklist

Title, Branch, and Commits

• PR title follows Conventional Commits: feat(minimax): support MiniMax-Text-01 and MiniMax-M2.
• Branch name follows <type>/xxx-yyyy-zzzz: feat/minimax-model-support.
• Each commit message follows Conventional Commits.
• This is a two-commit PR; each commit is focused and meaningful.
• No fixup!, squash!, or wip commits remain.
• Branch is based directly on InfiniLM-v0.2.9c, not the refactored main branch.

Scope and Design

• Changes are limited to MiniMax model support, tests, and related documentation.
• No debug prints or stray TODO/FIXME were found in the MiniMax source and test paths.
• No unrelated formatting churn is intended.
• No public API changes are introduced.

General Code Hygiene

• Added files end with a newline.
• No trailing whitespace or BOMs are intended.
• Comments and error messages are in English.
• Changed files have not yet been run through scripts/format.py on this branch.

C++ Specific

• Google C++ formatting and scripts/format.py verification are still pending.
• No raw new/delete were intentionally introduced.
• No changes or references to csrc/models/llama_legacy/.

Python Specific

• scripts/format.py verification is still pending.
• No changes or references to python/infinilm/auto_config.py.

Testing

• CPU operator test passed.
• CPU smoke test passed.
• CPU HuggingFace alignment test passed.
• NVIDIA model-level test is pending on the old architecture branch.
• Single request test was not run yet.
• Offline performance test was not run; no performance claim.
• Sanity benchmark test was not run yet.
• Service test was not run yet.

Build, CI, and Tooling

• The model was built and tested on CPU in the archived branch.
• CI has not been triggered for this branch yet.

Documentation

• MiniMax model documentation was added under docs/minimax/.

Security and Safety

• No secrets, tokens, private URLs, or personal hardware identifiers are intentionally included.
• No third-party code was added.
• No unsafe pointer arithmetic was intentionally introduced.

NVIDIA RTX 4090 / sm_89 Validation (2026-09-19)

Validated with:

• InfiniCore feat/minimax-lightning-attention at 7bcaa623
• InfiniLM feat/minimax-model-support at e23eba79
• RTX 4090, compute capability sm_89, driver 570.124.06, CUDA 12.8 (V12.8.61)
• PyTorch 2.6.0a0+ecf3bae40a.nv25.01 with CUDA 12.8, torch.cuda.is_available() == True
• Python 3.12.3, g++ 13.3.0, xmake 3.1.1+20260919

Results

Test	Device	Result	Error
MiniMax lightning attention operator	NVIDIA	Passed	decode: out 4.768e-07, state 2.384e-07; prefill: out 1.907e-06, state 4.768e-07
MiniMax smoke prefill/decode	NVIDIA	Passed	finite logits; state continuity 3.127381e-06
MiniMax vs HuggingFace, 1 expert, FP32	NVIDIA	Passed	prefill 5.257875e-04; decode 2.117679e-04
MiniMax vs HuggingFace, 4 experts, top-2, FP16	NVIDIA	Passed	prefill 7.324219e-04; decode 3.662109e-04

All tests are below the 1e-2 model-level tolerance. The 4-expert top-2 result exercises the CUDA Fused MoE path rather than the dense fallback.

Validation Commands

source /data/InfiniCore/scripts/set_env_linux.sh
cd /data/InfiniLM

MINIMAX_DEVICE=cuda python3 test/models/minimax/test_lightning_attention_op.py
MINIMAX_DEVICE=cuda python3 test/models/minimax/smoke_minimax.py
MINIMAX_DEVICE=cuda python3 test/models/minimax/test_minimax_vs_hf.py

MINIMAX_DEVICE=cuda \
MINIMAX_TORCH_DTYPE=float16 \
MINIMAX_NUM_LOCAL_EXPERTS=4 \
MINIMAX_NUM_EXPERTS_PER_TOK=2 \
python3 test/models/minimax/test_minimax_vs_hf.py

Fixes Included

• CUDA Lightning Attention now supports FP16 and BF16 in addition to FP32.
• MiniMax linear-attention state uses the model dtype instead of always allocating FP32.
• Sampling offsets are copied to host before host-side indexing on CUDA.
• HuggingFace MiniMax MoE weights are remapped from block_sparse_moe.experts.{i}.{w1,w2,w3} to moe.experts.{i}.{gate_proj,up_proj,down_proj} and block_sparse_moe.gate to moe.gate.
• Lightning-attention GPU tests use disjoint initial/final state rows to avoid cross-request races.

Server Logs

• /data/minimax-final-gpu-op.log
• /data/minimax-final-gpu-smoke.log
• /data/minimax-final-gpu-vs-hf.log
• /data/minimax-gpu-test-moe4-top2-fp16-final.log