[ROCm][AMD] GLM5.2/5.3 Performance Optimization on gfx950 / MI355X

This feature page tracks the peformance optimization of amd/GLM-5.2-MXFP4 / amd/GLM-5.3-Quark-MXFP4-AttnFP8 on gf950/MI355X The following list of PRs lands as stacked PRs on vllm-project/vllm main.

CC: @tjtanaa, @jhartika-amd, @maeehart, @nholmber

Nvidia equivalence

• [In review] Index Share feature (1% E2E TTFT improvement) - vllm-project/vllm#51309
• [In review] Remove aten fill functor in prefill MQA logits (1% E2E TTFT improvement) - vllm-project/vllm#51314

AITER kernel optimization:

• [In review] use AITER kernel for cp_gather_indexer_k_quant_cache and indexer_k_quant_and_cache and top_k_per_row_decode instead of triton - vllm-project/vllm#46172
• [In review] Sparse indexer kernel fusion for K normalization, Q/K RoPE, FP8 quantization, and K-cache write - indexer_qk_rope_quant_and_cache (1% E2E throughput improvement) - vllm-project/vllm#51315
• [In review] MLA fused kernel for rope, query concat, KV concat, fp8 cache write - fused_qk_rope_concat_and_cache_mla - vllm-project/vllm#47757
• [In review] Replace ATen gemm with tuned AITER gemm aiter.tuned_gemm.tgemm.mm (3% E2E throughput improvement) - vllm-project/vllm#50535
• [closed] fMoE tuning for gfx950 (5% E2E throughput improvement) - ROCm/aiter#4629
• [open in aiter] Dispatch to FlyDSL MQA logits in prefill flydsl_fp8_mqa_logits (1.5% E2E throughput improvement)

Additional optimization

• [ready] FP8 dtype hoist - vllm-project/vllm#53792

AITER dependencies that need to be merged

• FlyDSL paged MQA logits in decode - ROCm/aiter#4221
• FlyDSL MQA logits in prefill - ROCm/aiter#4538
• FlyDSL topk - ROCm/aiter#4355

Planned tasks

• [open in aiter] Dispatch to FlyDSL paged MQA logits in decode flydsl_fp8_paged_mqa_logits (1 week ETA, expected 1.5% E2E throughput improvement)
• MTP enablement:
  • support next_n > 2
• AITER kernel support
  • FlyDSL kernel for quick all reduce (INT4) - ROCm/aiter#4970​