[Kernel] Migrate scattered quantization kernels to sglang.kernels (RFC #29630, Phase 2.5, 1/7) (#30784)

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

benchmark/kernels/deepseek/benchmark_deepgemm_fp8_gemm.py

@@ -12,7 +12,7 @@ from vllm.model_executor.layers.quantization.utils.fp8_utils import (
 )
 
 from sglang.benchmark.bench_utils import run_bench
-from sglang.srt.layers.quantization.fp8_kernel import (
+from sglang.kernels.ops.quantization.fp8_kernel import (
     w8a8_block_fp8_matmul_deepgemm as w8a8_block_fp8_matmul,
 )
 

benchmark/kernels/deepseek/benchmark_deepgemm_fp8_gemm_blackwell.py

@@ -7,7 +7,7 @@ from deep_gemm import ceil_div
 from flashinfer.gemm import gemm_fp8_nt_groupwise
 
 from sglang.benchmark.bench_utils import run_bench
-from sglang.srt.layers.quantization.fp8_kernel import (
+from sglang.kernels.ops.quantization.fp8_kernel import (
     sglang_per_token_group_quant_fp8,
     w8a8_block_fp8_matmul_deepgemm,
 )

benchmark/kernels/flashinfer_allreduce_fusion/benchmark_fused_collective.py

@@ -28,6 +28,8 @@ from typing import Optional
 import torch  # type: ignore
 import torch.distributed as dist  # type: ignore
 
+from sglang.kernels.ops.quantization.fp8_kernel import fp8_dtype as SGLANG_FP8_DTYPE
+from sglang.kernels.ops.quantization.fp8_kernel import static_quant_fp8
 from sglang.srt.distributed import get_tp_group, tensor_model_parallel_all_reduce
 from sglang.srt.distributed.parallel_state import (
     cleanup_dist_env_and_memory,
@@ -36,8 +38,6 @@ from sglang.srt.distributed.parallel_state import (
     initialize_model_parallel,
 )
 from sglang.srt.layers.layernorm import RMSNorm  # noqa
-from sglang.srt.layers.quantization.fp8_kernel import fp8_dtype as SGLANG_FP8_DTYPE
-from sglang.srt.layers.quantization.fp8_kernel import static_quant_fp8
 
 try:
     from sgl_kernel import fused_add_rmsnorm as SGL_FUSED_ADD_RMS_NORM

benchmark/kernels/quantization/bench_int8_quant.py

@@ -5,7 +5,7 @@ import triton
 from vllm._custom_ops import scaled_int8_quant as vllm_scaled_int8_quant
 
 from sglang.benchmark.bench_utils import run_bench
-from sglang.srt.layers.quantization.int8_kernel import per_token_quant_int8
+from sglang.kernels.ops.quantization.int8_kernel import per_token_quant_int8
 
 
 @torch.compile(backend="inductor")

benchmark/kernels/quantization/tuning_block_wise_kernel.py

@@ -27,11 +27,11 @@ from tqdm import tqdm
 
 mp.set_start_method("spawn", force=True)
 
-from sglang.srt.layers.quantization.fp8_kernel import (
+from sglang.kernels.ops.quantization.fp8_kernel import (
     _w8a8_block_fp8_matmul,
     _w8a8_block_fp8_matmul_unrolledx4,
 )
-from sglang.srt.layers.quantization.int8_kernel import _w8a8_block_int8_matmul
+from sglang.kernels.ops.quantization.int8_kernel import _w8a8_block_int8_matmul
 from sglang.srt.utils import (
     get_device,
     get_device_core_count,
@@ -521,7 +521,9 @@ if __name__ == "__main__":
     parser.add_argument("--block-k", type=int, default=128)
     parser.add_argument("--batch-sizes", nargs="+", type=int, required=False)
     parser.add_argument(
-        "--save-path", type=str, default="python/sglang/srt/layers/quantization/configs"
+        "--save-path",
+        type=str,
+        default="python/sglang/kernels/ops/quantization/configs",
     )
     args = parser.parse_args()