[AMD] fix dsv4 indexer dtype dispatch on gfx950

[billishyahao]

Motivation

The DeepSeek-V4 indexer and the unified-KV paged-decode kernel selected their FP8 data type with is_hip(), which hardcodes torch.float8_e4m3fnuz for all AMD GPUs. This is only correct on CDNA3 (gfx942). This PR aligns the dtype so the dispatch is correct on gfx950 (and unchanged on gfx942 / NVIDIA).

Modifications

• python/sglang/srt/layers/attention/dsv4/indexer.py
  • Select FP8_DTYPE / FP8_MAX via is_fp8_fnuz() instead of is_hip().
  • Import is_fp8_fnuz from sglang.srt.layers.quantization.fp8_kernel; drop the
    now-unused is_hip import.
• python/sglang/srt/layers/attention/dsv4/unified_kv_kernels/paged_decode.py
  • Select _FP8_DTYPE via is_fp8_fnuz() instead of the hardcoded
    torch.float8_e4m3fnuz; add the import and update the storage comment.

Accuracy Tests

GSM8K on gfx950 (MI355) with DeepSeek-V4-Pro, tp=8, --kv-cache-dtype fp8_e4m3, --attention-backend dsv4:

lm_eval --model local-completions \
  --model_args model=/models/DeepSeek-V4-Pro,base_url=http://localhost:8000/v1/completions,num_concurrent=16,max_retries=3,tokenized_requests=False \
  --tasks gsm8k --num_fewshot 5

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.9507|±  |0.0060|
|     |       |strict-match    |     5|exact_match|↑  |0.9515|±  |0.0059|

Speed Tests and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

---

CI States

Latest PR Test (Base): ❌ Run #29009248782
Latest PR Test (Extra): ❌ Run #29009248687

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[HaiShaw]

@billishyahao cross repos, we use abs up to 224.0, not 240.0 as FP8_MAX for e4m3fnuz.

[billishyahao]

Get it! I re-checked the usage and FP8_MAX is invoked within indexer.py so I removed this part totally.