minimax fp8 index cache write

[ganyi1996ppo]

Motivation

Technical Details

Test Plan

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
ci:sglang	SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom	ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full	ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm	vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all	All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 4098 --add-label <label>

[Copilot]

Pull request overview

This PR updates the fused fused_qknorm_idxrqknorm sparse-insert path to support writing the index cache using the same cache element type as the KV caches (including FP8 / uint8-backed FP8 storage), and adjusts the Python perf test to allocate and validate the index cache accordingly.

Changes:

• Update the kernel signature and store path so index_cache is a cache_t* and uses the FP8 cache store helper (unit scale) in the FP8 path.
• Relax/adjust C++ input validation so FP8 insert mode accepts index_cache as FP8 or uint8 storage (and non-FP8 mode still requires dtype match with qkv).
• Update the op_test to optionally allocate index_cache with the cache dtype and to use an FP8 unit-scale reference for the per-token FP8 mode.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
op_tests/test_fused_qknorm_idxrqknorm.py	Allocates index cache with optional fp8/uint8 dtype and updates reference comparisons for fp8 index-cache writes.
csrc/kernels/fused_qknorm_idxrqknorm.cu	Switches index cache to use cache element type and stores via FP8-aware cache write helper; updates dtype checks and launch plumbing.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.