Commit 083b80ea authored by zhuwenwen's avatar zhuwenwen
Browse files

增加w8a8相关修改

parent 09428eec
......@@ -661,7 +661,7 @@ if _build_custom_ops():
ext_modules.append(CMakeExtension(name="vllm._C"))
package_data = {
"vllm": ["py.typed", "model_executor/layers/fused_moe/configs/*.json", "benchmarks/*.py","model_executor/layers/quantization/configs/*.json"]
"vllm": ["py.typed", "model_executor/layers/fused_moe/configs/*.json", "benchmarks/*.py","model_executor/layers/quantization/configs/w8a8/*.json"]
}
if _no_device():
......
......@@ -923,7 +923,7 @@ def triton_int8_gemm_helper(m: int,
out_dtype: Type[torch.dtype] = torch.float16,
device: str = "cuda",
best_config:Optional[list] = None):
return quant_tools.triton_int8_gemm_helper(m,n,k,per_token_act_quant,per_out_channel_weight_quant,use_bias,out_dtype,device,est_config)
return quant_tools.triton_int8_gemm_helper(m,n,k,per_token_act_quant,per_out_channel_weight_quant,use_bias,out_dtype,device,best_config)
def cutlass_scaled_mm_azp(a: torch.Tensor,
......
......@@ -226,7 +226,7 @@ class AWQLinearMethod(LinearMethodBase):
deqweight=ops.dequant_w4_gemm_colmajor( # shape[n, k/8] ---> [n,k]
qweight,
zeros_and_scales,
k,
k+padding_group*self.quant_config.group_size,
n,
self.quant_config.group_size)
out=F.linear(reshaped_x, deqweight[:,0:k])
......
......@@ -24,7 +24,7 @@ class CompressedTensorsW8A8Int8(CompressedTensorsScheme):
input_symmetric: bool):
self.strategy = strategy
self.is_static_input_scheme = is_static_input_scheme
self.w8a8_strategy=int(os.getenv('W8A8_SUPPORT_METHODS', '0'))
self.w8a8_strategy=int(os.getenv('W8A8_SUPPORT_METHODS', '1'))
self.input_symmetric = input_symmetric
@classmethod
......
{
"12288_4096": {
"1": {
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"2": {
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"3": {
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"4": {
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
......@@ -36,17 +36,17 @@
"num_stages": 1,
"num_warps": 8
},
"5": {
"BLOCK_SIZE_M": 16,
"36": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"6": {
"BLOCK_SIZE_M": 16,
"40": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
......@@ -54,17 +54,17 @@
"num_stages": 1,
"num_warps": 8
},
"7": {
"BLOCK_SIZE_M": 16,
"44": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"8": {
"BLOCK_SIZE_M": 16,
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
......@@ -72,8 +72,8 @@
"num_stages": 1,
"num_warps": 8
},
"9": {
"BLOCK_SIZE_M": 16,
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
......@@ -81,17 +81,17 @@
"num_stages": 1,
"num_warps": 8
},
"10": {
"BLOCK_SIZE_M": 16,
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"11": {
"BLOCK_SIZE_M": 16,
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
......@@ -99,35 +99,53 @@
"num_stages": 1,
"num_warps": 8
},
"12": {
"BLOCK_SIZE_M": 16,
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"13": {
"BLOCK_SIZE_M": 32,
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"14": {
"BLOCK_SIZE_M": 16,
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"15": {
"BLOCK_SIZE_M": 16,
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
......@@ -135,17 +153,17 @@
"num_stages": 1,
"num_warps": 8
},
"16": {
"BLOCK_SIZE_M": 16,
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"17": {
"BLOCK_SIZE_M": 16,
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
......@@ -153,61 +171,79 @@
"num_stages": 1,
"num_warps": 8
},
"18": {
"BLOCK_SIZE_M": 16,
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
"num_warps": 4
},
"19": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
"num_warps": 4
},
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
"num_warps": 4
},
"21": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
"num_warps": 4
},
"22": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_stages": 2,
"num_warps": 8
},
"23": {
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_stages": 2,
"num_warps": 8
},
"24": {
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
......@@ -216,16 +252,16 @@
"num_stages": 1,
"num_warps": 8
},
"25": {
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"26": {
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
......@@ -234,52 +270,52 @@
"num_stages": 1,
"num_warps": 8
},
"27": {
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_stages": 2,
"num_warps": 8
},
"28": {
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_stages": 2,
"num_warps": 8
},
"29": {
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"30": {
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_stages": 2,
"num_warps": 8
},
"31": {
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"32": {
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
......@@ -288,23 +324,41 @@
"num_stages": 1,
"num_warps": 8
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
"num_stages": 2,
"num_warps": 8
},
"128": {
"BLOCK_SIZE_M": 64,
"14": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"15": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
"num_warps": 8
},
"256": {
"BLOCK_SIZE_M": 128,
......@@ -330,7 +384,7 @@
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_stages": 2,
"num_warps": 4
},
"2048": {
......
{
"1280_8192": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"20": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"28": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"40": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"44": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"72": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 4
},
"80": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 4
},
"88": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 4
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 2,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"112": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"120": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"128": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 2
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 2,
"num_warps": 8
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 2,
"num_warps": 8
},
"512": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 2,
"num_warps": 8
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 2,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 8
}
}
}
\ No newline at end of file
{
"13824_5120": {
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"2": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"3": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"4": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"5": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"6": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"7": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"8": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"9": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"10": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"11": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"12": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"14": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"15": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"28": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"40": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"44": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
}
}
}
\ No newline at end of file
{
"14336_8192": {
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"2": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"3": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"4": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"5": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"6": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"7": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"8": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"9": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"10": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"11": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"12": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"13": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"14": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"15": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"17": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"40": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
{
"15360_5120": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"28": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"40": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"52": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"56": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"60": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
}
}
}
\ No newline at end of file
{
"22016_4096": {
"1": {
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 2
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"2": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"3": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4": {
"36": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
"num_warps": 4
},
"5": {
"40": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 256,
......@@ -45,34 +72,34 @@
"num_stages": 1,
"num_warps": 8
},
"6": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
"num_stages": 0,
"num_warps": 4
},
"7": {
"56": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"8": {
"60": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"9": {
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 256,
......@@ -81,245 +108,272 @@
"num_stages": 1,
"num_warps": 8
},
"10": {
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"11": {
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"12": {
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"13": {
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"14": {
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"15": {
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"16": {
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"17": {
"BLOCK_SIZE_M": 16,
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_stages": 1,
"num_warps": 8
},
"18": {
"BLOCK_SIZE_M": 16,
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_stages": 1,
"num_warps": 8
},
"19": {
"BLOCK_SIZE_M": 16,
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_stages": 1,
"num_warps": 8
},
"20": {
"BLOCK_SIZE_M": 16,
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_stages": 1,
"num_warps": 8
},
"21": {
"BLOCK_SIZE_M": 16,
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_stages": 1,
"num_warps": 8
},
"22": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
"num_warps": 4
},
"23": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"2": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"3": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
"num_warps": 4
},
"25": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"4": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"5": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
"num_warps": 4
},
"26": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"6": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
"num_warps": 4
},
"27": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"7": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
"num_warps": 4
},
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"8": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
"num_warps": 4
},
"29": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"9": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
"num_warps": 4
},
"30": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"10": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
"num_warps": 4
},
"31": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"11": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
"num_warps": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"12": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
"num_warps": 4
},
"64": {
"13": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"14": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
"num_stages": 0,
"num_warps": 4
},
"15": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_stages": 2,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
......
{
"2560_8192": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"17": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"20": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"28": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"40": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"136": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"144": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"152": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"160": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"16384": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
{
"27648_5120": {
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"2": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"40": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
{
"28672_4096": {
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"2": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"3": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"5": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"6": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"7": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"9": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"10": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"11": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"12": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"14": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"15": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"40": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"44": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"256": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
}
}
}
\ No newline at end of file
{
"28672_8192": {
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"2": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"3": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"5": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"6": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"7": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"9": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"10": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"11": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"12": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"13": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"14": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"15": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"40": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"44": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"52": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"56": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"60": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
{
"32000_4096": {
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"2": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"3": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"5": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"6": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"7": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"8": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"9": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"10": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"11": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"12": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"14": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"15": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"40": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"44": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"256": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
}
}
}
\ No newline at end of file
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment