[AMD-yanfeiwang]

Motivation

DeepSeek V4 C4 computes 2048-wide main-compressor and 512-wide indexer projections from the same hidden state. This PR, stacked on #38919, combines them into one 2560-wide GEMM and splits the result into zero-copy row-strided views. PR2-only diff.

Modifications

• Build fused weight aliases after loading and rebuild them after state assignment or device/dtype conversion.
• Run one GEMM and pass the two views into the existing compressor tails.
• Gate the ROCm path behind default-off SGLANG_OPT_DSV4_C4_COMPRESSOR_GEMM_FUSION=1.
• Fall back for IPC weight cache, offload, prefill CP, unsupported weights, and C128.

Accuracy Tests

• Fusion unit tests: 8 passed.
• C4 v2 tests from [ROCm][DSV4] Support row-strided inputs in C4 v2 compression #38919: 37 passed.
• IPC weight-cache protocol: 33 passed.
• DeepSeek-V4-Pro TP8 off/on/off generated the same 128 greedy token IDs and text. Strict cross-restart logprob parity was inconclusive because the two unfused controls also differed.

Speed Tests and Profiling

The projection microbenchmark improved 11/12 graph buckets, with a 10.29% median latency reduction.

DeepSeek-V4-Pro TP8 serving on 8x MI355X/gfx950 used one fresh-container unfused-A/fused/unfused-B sandwich. C1/C4/C8/C16 all used random lengths in the 0.8-1.0x range with 8K/1K maxima, seeds 1001/1002/1003, 10 * concurrency measured requests, and an explicit 2 * concurrency warmup before every sample. Each arm also ran two full-shape priming sweeps. Radix cache, HiCache, and LMCache were disabled. All 36 measured windows were free of serving-time compilation, and request/token workloads matched exactly across arms.

C	A / fused / B median tok/s	Matched-seed throughput delta	Matched-seed TPOT delta	Max A/B drift (throughput / TPOT)	Gate
1	83.38 / 84.98 / 85.67	+0.36% (+2.51/-0.60/+0.36%)	+0.19%	3.11% / 3.29%	Fail
4	255.31 / 254.98 / 253.69	-0.01% (-1.00/-0.01/+1.65%)	+0.08%	2.94% / 2.87%	Fail
8	424.40 / 425.89 / 417.67	+1.64% (+2.26/-1.07/+1.64%)	+1.77%	2.17% / 2.05%	Fail
16	625.61 / 652.85 / 638.72	+3.19% (+2.58/+4.57/+3.19%)	+2.99%	3.06% / 3.07%	Pass

The strict per-workload gate requires all three throughput and TPOT deltas to be positive, at least +1% median improvement for both metrics, and at most 5% A/B control drift per seed. C16 passes and supports a stable gain for this exact service workload. C1/C4 are neutral; C8 has a positive median but one negative seed. The feature remains default-off because API concurrency is not a reliable local-GEMM shape gate and the benefit is not uniform across workloads.

Checklist

• Feature is ROCm-only and disabled by default.
• Unsupported configurations fail closed to the existing path.

---

CI States

Latest PR Test (Base): ❌ Run #36813359346
Latest PR Test (Extra): ❌ Run #36813359178
Latest PR Test (AMD ROCm 10): ❌ Run #36813359504