fused rmsnorm quant fp8 work

[yctseng0211]

Motivation

Fusion of rmsnorm + group scale FP8 quantization (group_size=128)
This PR use the triton kernel from main branch of Rocm/aiter : ROCm/aiter#1148

w/o fusion

w/ fusion

Modifications

deepseek.py

• LayerCommunicator (for hidden states)
• DeepseekV2AttentionMLA (for q, k_)

f8.py
fp8_utils.py

• aiter_w8a8_block_fp8_linear: skip quantization here when x is tuple

Accuracy Tests

python3 -m sglang.launch_server
--model-path /DeepSeek-R1-0528/
--quantization fp8
--tensor-parallel-size 8
--trust-remote-code
--chunked-prefill-size 131072
--host 0.0.0.0
--port 8008
--log-requests
--disable-radix-cache
--mem-fraction-static 0.95 

python3 benchmark/gsm8k/bench_sglang.py --num-questions 2000 --parallel 2000 --port 8008

w/. fused_rms_quant at LayerCommunicator and DeepseekV2AttentionMLA
Accuracy: 0.961
Invalid: 0.000
Latency: 55.205 s
Output throughput: 2412.560 token/s

w/. fused_rms_quant at LayerCommunicator
Accuracy: 0.960
Invalid: 0.000
Latency: 55.688 s
Output throughput: 2366.073 token/s

w/o fused_rms_quant
Accuracy: 0.957
Invalid: 0.000
Latency: 57.499 s
Output throughput: 2302.590 token/s

Benchmarking and Profiling

docker image : docker.io/rocm/sgl-dev:v0.5.3.post2-rocm700-mi35x-20251016
Model : DeepSeek-R1-0528/

Server Mode

Concurrency	/DeepSeek-R1-0528/ docker.io/rocm/sgl-dev:v0.5.3.post2-rocm700-mi35x-20251016				wo fused / w fused
	MI355 SGLang + spec		MI355 SGLang + spec + rms quant fused		Latency (%)	Throughput (%)
	Latency (ms)	Throughput (tok/s)	Latency (ms)	Throughput (tok/s)
1	6214.4	614.0	6176.89	647.3	101%	105%
2	6577.2	1179.1	6481.13	1153.1	101%	98%
4	7595.5	2034.4	7271.11	2029.8	104%	100%
8	8407.2	3522.0	8338.99	3645.4	101%	104%
16	10511.4	5822.1	10519.08	5755.3	100%	99%
32	13928.9	8739.0	13749.6	8856.5	101%	101%
64	23414.9	10482.2	22691.04	10724.9	103%	102%

Offline Mode

Batch Size	Input Size	Output Size	Prefill Latency (%)	Decode Latency (%)
1	256	2048	82%	103%
2	256	2048	81%	103%
4	256	2048	69%	103%
8	256	2048	101%	102%
16	256	2048	101%	101%
32	256	2048	101%	100%
64	256	2048	101%	101%
128	256	2048	101%	100%
256	256	2048	101%	102%
512	256	2048	101%	101%

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.