Fuse the DSA (V3.2, GLM-5.x) indexer Q/K paths into single kernels

[b8zhong]

Note: to completely revert the affects of this PR, we can set SGLANG_DISABLE_DSA_INDEXER_FUSION=1

Changes

1. Fuse the key projection and the head-gate projection into a single matmul
   instead of two.

2. Fuse the query path — rope, optional Hadamard, fp8 quantization, and
   head-gate scaling — into one kernel instead of four separate ops.

3. Fuse the key path — normalization, rope, optional Hadamard, fp8 quantization,
   and the index cache write — into one kernel instead of four. The previous
   split path is kept as a fallback for the cases that can't use the fused store
   (non-CUDA / NPU, HIP).

4. Drop the Hadamard rotation for V3.2. It cancels out of the indexer scores, so
   removing it doesn't change results — it only affected quantization accuracy,
   and V3.2 now quantizes the un-rotated activations directly. It stays available
   for the paths that still want it.

5. Run the key and query paths concurrently. Previously the cache store sat on
   the critical path after the key work; folding the store into the key kernel
   lets the entire key side overlap the query side.

Note: for the fusion, it seems vLLM has adopt a similar approach: vllm-project/vllm#38928

Checkpoint / weight loading

The fused key+head-gate matmul is bf16, so the loader has to land the key
projection as bf16. How it gets there depends on the checkpoint format (the bf16
result is identical either way):

Checkpoint	indexer key projection	loader behavior
deepseek-ai/DeepSeek-V3.2 (FP8)	block-fp8 + scale	dequantize to bf16 at load
zai-org/GLM-5-FP8	block-fp8 + scale	dequantize to bf16 at load
nvidia/DeepSeek-V3.2-NVFP4	already bf16	straight copy
nvidia/GLM-5.1-NVFP4	already bf16	straight copy

The NVFP4 checkpoint authors intentionally left the (small) indexer out of
low-bit quantization and pre-dequantized its key projection to bf16, verified it
matches the FP8 checkpoint dequantized exactly. The head-gate projection is bf16
in all of them.

FP4:

Before:

After:

FP8:

Before:

After:

Accuracy Tests

GPQA:

nohup env CUDA_VISIBLE_DEVICES=4,5,6,7 python3 -m sglang.launch_server \
  --model-path nvidia/GLM-5.1-NVFP4 \
  --tp 4 \
  --ep 4 \
  --trust-remote-code \
  --reasoning-parser glm45 \
  --tool-call-parser glm47 \
  --port 30001 \
  --max-running-requests 256 > glm_server.log 2>&1 &

sglang serve \
  --model-path zai-org/GLM-5.2-FP8 \
  --tp 8 \
  --ep 8 \
  --trust-remote-code \
  --reasoning-parser glm45 \
  --tool-call-parser glm47 \
  --max-running-requests 256

nohup python3 -m sglang.test.run_eval --port 30020 --eval-name gpqa --num-examples 198 --max-tokens 128000 --repeat 8 --top-p 0.95 --temperature 1.0 --thinking-mode glm-45 > eval.log 2>&1 &

FP4:

Before:

Repeat: 8, mean: 0.854
Scores: ['0.838', '0.848', '0.869', '0.838', '0.854', '0.854', '0.864', '0.869']

After:

Repeat: 8, mean: 0.859
Scores: ['0.843', '0.828', '0.874', '0.864', '0.864', '0.848', '0.884', '0.864']

FP8:

Before:

Repeat: 8, mean: 0.856
Scores: ['0.869', '0.838', '0.889', '0.828', '0.838', '0.864', '0.864', '0.859']

After:

Repeat: 8, mean: 0.856
Scores: ['0.859', '0.859', '0.843', '0.843', '0.848', '0.859', '0.848', '0.889']

Speed Tests and Profiling

BS = 1: 97.72 -> 107.77 TPS with 0 context (about 10%), with above command on B300.

For B300 + GLM FP8 checkpoint and BS = 128, around 3212.05 -> 3418.63 TPS in decode

---

CI States

Latest PR Test (Base): ✅ Run #28271629599
Latest PR Test (Extra): ✅ Run #28302367849

[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!