[AMD-yanfeiwang]

Motivation

On the fused qk-norm/rope path (gfx95/gfx1250), MQALayer._forward_prepare already quantizes the normed q_lora to fp8 group-128 (_fused_rmsnorm_fp8_quant) for the attention wq_b. The C4 indexer's wq_b takes the same q_lora and quantizes it again with a standalone dynamic_per_group_scaled_quant. For DeepSeek-V4-Pro that adds 30 extra launches per decode/verify step.

Modifications

• _forward_prepare passes the existing quant (q_lora_quant) through C4Indexer.forward -> forward_c4_indexer -> compute_q, and the indexer wq_b consumes it directly.
• It is dropped (set to None) when the indexer trims q_lora to fewer queries. The non-fused path, the multi-stream path, NPU and the low-ratio arms are unchanged. forward_c4_indexer only receives the kwarg when it is set.

Accuracy Tests

• The quantized input is the same tensor the attention wq_b already uses, so the indexer wq_b input is identical to its own quant of q_lora.
• DeepSeek-V4-Pro TP8, stock image lmsysorg/sglang-rocm:v0.5.20-rocm724-mi35x-20260927 + this PR only: gsm8k 200q 0.965. A capture-time print confirmed that the reuse path runs on all 8 ranks.
• On the stack below, DSpark b6 gsm8k 1319q over 2 boots: before 0.945 / 0.939, after 0.945 / 0.948. Accept length 4.36 -> 4.37.

Speed Tests and Profiling

Kernel: each removed dynamic_per_group_scaled_quant launch shows as 4.3 us in the graph-mode trace (about 3.8 us after subtracting profiler overhead). With 30 launches per step, the trace estimate was about 0.11-0.13 ms per step. The measured E2E saving is 0.06-0.09 ms per step, which is about 2-3 us per launch.

E2E, DeepSeek-V4-Pro fp8, TP8 MI355X, image lmsysorg/sglang-rocm:v0.5.20-rocm724-mi35x-20260927, 8k in / 1k out. Measured on our DSv4 decode stack: #41383 (FlyDSL fused decode, with ROCm/aiter#5882 and the aiter fp32-out HGEMM fix), the DSpark verify ring write, #38919 + #38920 (env on), the MoE small-batch sort and the fp32 GEMM aiter route. Before = that stack; after = stack + this PR. It was developed there; the diff on main is the same minus the #38920 kv_score_input plumbing. 2 boots x 3 reps, paired:

cell	step time before -> after (ms)	delta mean [min, max] (ms)	steady decode tok/s
plain c8	11.87 -> 11.79	-0.65% -0.077 [-0.100, -0.043]	674 -> 678
plain c32	15.60 -> 15.54	-0.39% -0.061 [-0.087, -0.035]	2053 -> 2062
DSpark b6 c2	14.85 -> 14.76	-0.60% -0.089 [-0.113, -0.078]	859 -> 864
DSpark b6 c4	17.36 -> 17.30	-0.33% -0.058 [-0.135, -0.021]	1407 -> 1417
DSpark b6 c16 (1 boot)	26.26 -> 26.27	+0.06% (noise)	-
DSpark b6 c64 (1 boot)	45.13 -> 45.05	-0.18% (noise)	-

All 24 two-boot pairs are faster.

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.

---

CI States

Latest PR Test (Base): ❌ Run #38719443823
Latest PR Test (Extra): ❌ Run #38719443685
Latest PR Test (AMD ROCm 10): ❌ Run #38719443733

[HaiShaw]

None none q_lora_quant is only set at HIP/gfx95/gfx1250