feat(rocm): integrate packed GR read for Qwen3.8 Flash Next TP2 and TP4

[apinge]

Use flydsl for decode (batch 1-32), fallback to torch.compile for others. Introduce preshuffle weight to support flydsl gemm operations.

Add SGLANG_GR_READ_FLYDSL (default on) to select the integration or the original path at model construction.

Motivation

Modifications

1. in packed_weights apply preshuffling to both down gemm and up gemm weights via aiter.ops.shuffle.shuffle_weight For up gemm weights, perform an extra hc/hidden dimension conversion ( up_interleaved[h*4+c,r] = w_up[c*2560+h,r])before the preshuffle step.

Accuracy Tests

 python -m sglang.test.run_eval \
  --eval-name gsm8k --num-examples 1319 --max-tokens 16384 --port 9080
[aiter] WARNING: NUMA balancing is enabled, which may cause errors. It is recommended to disable NUMA balancing by running "sudo sh -c 'echo 0 > /proc/sys/kernel/numa_balancing'" for more details: https://rocm.docs.amd.com/en/latest/how-to/system-optimization/mi300x.html#disable-numa-auto-balancing
[aiter] import [module_aiter_core] under /opt/aiter/aiter/jit/module_aiter_core.so
Downloading from https://raw.githubusercontent.com/openai/grade-school-math/master/grade_school_math/data/test.jsonl to /tmp/test.jsonl
/tmp/test.jsonl: 732kB [00:00, 3.87MB/s]                                                                                                                                                                            
ChatCompletionSampler initialized with self.system_message=None self.temperature=0.0 self.max_tokens=16384 self.reasoning_effort=None self.extra_body=None self.stop=None self.record_meta_info=True
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1314/1314 [21:23<00:00,  1.02it/s]
Total latency: 1283.525 s
Score: 0.979
Output throughput: 490.490 token/s
[METRIC] gsm8k_score=0.9794520547945206 labels={"model": "Qwen/Qwen3.8-Flash-Next-PTPC-FP8", "eval": "gsm8k"}
[METRIC] gsm8k_latency=1283.5254440808203 labels={"model": "Qwen/Qwen3.8-Flash-Next-PTPC-FP8", "eval": "gsm8k"}

Speed Tests and Profiling

profiled with concurrency 1

Before

After (This PR)

Comparison with triton 3 stage

From #35

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.

[sammysun0711]

One dummy question, for a server running with BS 1-64 in case prefix cache enabled.
We have flydsl GR read and original torch compile kernel, do we trigger weight preshuffle for BS 1-32 via pack_weights and unpack_weight for BS > 32 during runtime?

[sammysun0711]

Let's remove markdown after PR ready. SGLANG_GR_READ_FLYDSL=1 should be mentioned in PR for enable this opt.