From bdfb02674fc0eaa2b9023420b02a8d4e23a69957 Mon Sep 17 00:00:00 2001 From: pmanczak Date: Fri, 2 Oct 2026 13:56:11 +0200 Subject: [PATCH] [XPU][Kernel] Tune Triton W8A8 block-FP8 GEMM for Intel B70 (#56063) Signed-off-by: Pawel Manczak <203671809+pmanczak@users.noreply.github.com> Co-authored-by: Pawel Manczak <203671809+pmanczak@users.noreply.github.com> --- .../kernels/benchmark_w8a8_block_fp8.py | 51 ++++++++++-- ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ ...,dtype=fp8_w8a8,block_shape=[128,128].json | 82 +++++++++++++++++++ 27 files changed, 2174 insertions(+), 9 deletions(-) create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=2048,K=11008,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=2048,K=2048,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=2112,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=22016,K=2048,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=24576,K=4096,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=28672,K=4096,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=3072,K=1536,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=3072,K=2048,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=3072,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=4096,K=12288,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=4096,K=14336,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=4096,K=4096,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=4096,K=512,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=4608,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=512,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=576,K=2048,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=6144,K=4096,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2048,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2304,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json create mode 100644 vllm/model_executor/layers/quantization/utils/configs/N=7168,K=256,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json diff --git a/benchmarks/kernels/benchmark_w8a8_block_fp8.py b/benchmarks/kernels/benchmark_w8a8_block_fp8.py index 3ad7ec8887..f543ee520b 100644 --- a/benchmarks/kernels/benchmark_w8a8_block_fp8.py +++ b/benchmarks/kernels/benchmark_w8a8_block_fp8.py @@ -23,9 +23,11 @@ from vllm.utils.platform_utils import get_device_name_as_file_name mp.set_start_method("spawn", force=True) -assert current_platform.is_cuda() or current_platform.is_rocm(), ( - "Only support tune w8a8 block fp8 kernel on CUDA/ROCm device." -) +assert ( + current_platform.is_cuda() + or current_platform.is_rocm() + or current_platform.is_xpu() +), "Only support tune w8a8 block fp8 kernel on CUDA/ROCm/XPU device." DTYPE_MAP = { "float32": torch.float32, @@ -138,6 +140,29 @@ def get_configs_compute_bound(): return configs +def get_xpu_tuning_space(): + # NOTE: curated subset of get_configs_compute_bound(); the full grid is + # too slow to tune on XPU due to per-config Triton JIT compile time. + configs = [] + for num_stages in [2, 3]: + for block_m in [32, 64, 128]: + for block_k in [64, 128]: + for block_n in [128, 256]: + for num_warps in [4, 8, 16, 32]: + for group_size in [1, 32]: + configs.append( + { + "BLOCK_SIZE_M": block_m, + "BLOCK_SIZE_N": block_n, + "BLOCK_SIZE_K": block_k, + "GROUP_SIZE_M": group_size, + "num_warps": num_warps, + "num_stages": num_stages, + } + ) + return configs + + def get_weight_shapes(tp_size): # NOTE(HandH1998): The weight shapes only works for DeepSeek-V3. # Modify them, if you tune for another different model. @@ -208,13 +233,14 @@ def tune(M, N, K, block_size, out_dtype, search_space, input_type): fp8_info = torch.finfo(torch.float8_e4m3fn) fp8_max, fp8_min = fp8_info.max, fp8_info.min + device = current_platform.device_type A_fp32 = ( - (torch.rand(M, K, dtype=torch.float32, device="cuda") - 0.5) * 2 * fp8_max + (torch.rand(M, K, dtype=torch.float32, device=device) - 0.5) * 2 * fp8_max ) A = A_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn) B_fp32 = ( - (torch.rand(N, K, dtype=torch.float32, device="cuda") - 0.5) * 2 * fp8_max + (torch.rand(N, K, dtype=torch.float32, device=device) - 0.5) * 2 * fp8_max ) B = B_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn) else: @@ -224,9 +250,9 @@ def tune(M, N, K, block_size, out_dtype, search_space, input_type): n_tiles = (N + block_n - 1) // block_n k_tiles = (K + block_k - 1) // block_k - As = torch.rand(M, k_tiles, dtype=torch.float32, device="cuda") * factor_for_scale + As = torch.rand(M, k_tiles, dtype=torch.float32, device=device) * factor_for_scale Bs = ( - torch.rand(n_tiles, k_tiles, dtype=torch.float32, device="cuda") + torch.rand(n_tiles, k_tiles, dtype=torch.float32, device=device) * factor_for_scale ) @@ -297,7 +323,11 @@ def tune_on_gpu(args_dict): save_path = args.save_path input_type = args.input_type - search_space = get_configs_compute_bound() + search_space = ( + get_xpu_tuning_space() + if current_platform.is_xpu() + else get_configs_compute_bound() + ) search_space = [ config for config in search_space if block_k % config["BLOCK_SIZE_K"] == 0 ] @@ -342,7 +372,10 @@ def main(args): raise RuntimeError("No GPU available for tuning") print(f"Found {num_gpus} GPUs for parallel tuning") - torch.cuda.init() + if current_platform.is_xpu(): + torch.xpu.init() + else: + torch.cuda.init() if args.batch_size is None: batch_sizes = [ diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..85a2aa9d83 --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=2048,K=11008,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=2048,K=11008,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..a8bc891dd5 --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=2048,K=11008,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 2 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=2048,K=2048,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=2048,K=2048,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..072db2cbc7 --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=2048,K=2048,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=2112,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=2112,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..8b535371f0 --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=2112,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 2 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=22016,K=2048,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=22016,K=2048,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..323db388eb --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=22016,K=2048,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 2 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=24576,K=4096,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=24576,K=4096,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..5ce80c0316 --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=24576,K=4096,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 3 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..e1a348c85d --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 2 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=28672,K=4096,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=28672,K=4096,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..eeada919b2 --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=28672,K=4096,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 3 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=3072,K=1536,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=3072,K=1536,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..9d16cde36f --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=3072,K=1536,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=3072,K=2048,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=3072,K=2048,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..5ed9e9d441 --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=3072,K=2048,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=3072,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=3072,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..d35330be1b --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=3072,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 2 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..dc895d5a29 --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=4096,K=12288,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=4096,K=12288,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..ff4536c455 --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=4096,K=12288,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 2 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=4096,K=14336,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=4096,K=14336,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..0ef06c44e6 --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=4096,K=14336,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=4096,K=4096,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=4096,K=4096,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..f8b9e8c696 --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=4096,K=4096,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 3 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=4096,K=512,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=4096,K=512,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..fd0318f8b5 --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=4096,K=512,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=4608,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=4608,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..0087fa0de0 --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=4608,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 2 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=512,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=512,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..d5af629a01 --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=512,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=576,K=2048,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=576,K=2048,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..bafaf2ad8b --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=576,K=2048,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..998adaa968 --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=6144,K=4096,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=6144,K=4096,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..6153e4b540 --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=6144,K=4096,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 2 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..a70cd56ba4 --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 2 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..8cc0253e39 --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 3 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2048,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2048,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..793df30860 --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2048,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2304,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2304,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..12246c9700 --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2304,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 32, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_stages": 3 + } +} diff --git a/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=256,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json b/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=256,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000..d529d1631a --- /dev/null +++ b/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=256,device_name=Intel(R)_Arc(TM)_Pro_B70_Graphics,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,82 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 16, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 32, + "num_stages": 3 + } +}