[Triton] DS FP4/FP8 Triton fusion and GEMM optimization

[k50112113]

This PR is co-authored by @k50112113, @omuhamma (#61) and @farlukas (#116)

This PR provides Triton fusion/GEMM optimizations for DS FP4 and FP8,
please use the following AITER branch for testing for now as some of the PRs are yet to be merged to AITER main:
https://github.com/ROCm/aiter/tree/shaoclee/atom_triton_tmp_0106

The required AITER PRs include:

1. [Triton] Triton A16WFP4 GEMM prequant aiter#1777
2. [Triton] Triton a16w8 gemm preshuffle aiter#1778
3. [Triton] Add Fused GEMM A8W8 + Split + Concat Triton Kernel aiter#1553 (review)

To activate the optimizations on ATOM, the following env variables are required:

# for concurrency > 4, use AR + RMS_Quant + GEMM optimizations:
export ATOM_USE_TRITON_GEMM=1
# note: ATOM_ENABLE_DS_INPUT_RMSNORM_QUANT_FUSION is turned on automictically when ATOM_USE_TRITON_GEMM is on

# for concurrency = 4, use AR_RMS + Quant_GEMM optimizations:
export ATOM_USE_TRITON_GEMM=1
export ATOM_ENABLE_DS_INPUT_RMSNORM_QUANT_FUSION=0

The following command along with the above env var are used to derive e2e performance results:

# for DS FP8
python -m atom.entrypoints.openai_server \
    --model /data/deepseek-ai/DeepSeek-R1-0528/ \
    -tp 8 \
    --block-size 1 \
    --server-port 8989 2>&1 | tee server.out

# for DS FP4
export ATOM_USE_TRITON_MXFP4_BMM=1
export AMDGCN_USE_BUFFER_OPS=1
python -m atom.entrypoints.openai_server \
    --model /data/DeepSeek-R1-0528-MXFP4-Preview \
    -tp 8 \
    --block-size 16 \
    --kv_cache_dtype fp8 \
    --server-port 8989 \
    2>&1 | tee server.out

For client command:

MODEL=<DS FP4 or FP8 model paths>
ISL=3500
OSL=1500
PORT=8989
for CONC in 4 256 128 64 32 16 8; do
    RESULT_FILENAME=${ISL}_${OSL}_${CONC}
    python /root/ATOM/atom/benchmarks/benchmark_serving.py \
        --model=$MODEL --backend=vllm --base-url=http://localhost:$PORT \
        --dataset-name=random \
        --random-input-len=$ISL --random-output-len=$OSL \
        --random-range-ratio 1.0 \
        --num-prompts=$(( $CONC * 8 )) \
        --max-concurrency=$CONC \
        --request-rate=inf --ignore-eos \
        --save-result --percentile-metrics="ttft,tpot,itl,e2el" \
        --result-dir=./ --result-filename=$RESULT_FILENAME.json 2>&1 | tee -a ${RESULT_FILENAME}.log
done

DS FP8 performance comparisons and uplift

DS FP4 performance comparisons and uplift

[ChuanLi1101]

Overall LGTM, approved for benchmark testing.

[k50112113]

Addressed all the comments

[ChuanLi1101]

Approved due to comments has addressed.