sglang 0.4.2.post3__py3-none-any.whl → 0.4.3__py3-none-any.whl

sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json

@@ -0,0 +1,175 @@
+{
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 0,
+        "waves_per_eu": 4,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 0,
+        "waves_per_eu": 1,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 1
+    },
+    "16": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 0,
+        "waves_per_eu": 2,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "32": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 0,
+        "waves_per_eu": 1,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "64": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 0,
+        "waves_per_eu": 2,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 0,
+        "waves_per_eu": 1,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 1
+    },
+    "256": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 0,
+        "waves_per_eu": 2,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 0,
+        "waves_per_eu": 4,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 0,
+        "waves_per_eu": 2,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 0,
+        "waves_per_eu": 2,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "8192": {
+        "BLOCK_SIZE_M": 256,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 0,
+        "waves_per_eu": 2,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 1
+    },
+    "16384": {
+        "BLOCK_SIZE_M": 256,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 0,
+        "waves_per_eu": 1,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 1
+    },
+    "32768": {
+        "BLOCK_SIZE_M": 256,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 0,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 1
+    },
+    "65536": {
+        "BLOCK_SIZE_M": 256,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 0,
+        "waves_per_eu": 1,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 1
+    },
+    "131072": {
+        "BLOCK_SIZE_M": 256,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 0,
+        "waves_per_eu": 2,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    }
+}

sglang/srt/layers/moe/fused_moe_triton/fused_moe.py

@@ -18,7 +18,7 @@ from sglang.srt.layers.quantization.fp8_kernel import per_token_group_quant_fp8
 from sglang.srt.utils import direct_register_custom_op, get_device_name, is_hip
 
 is_hip_flag = is_hip()
-from sgl_kernel import moe_align_block_size as sgl_moe_align_block_size
+
 
 logger = logging.getLogger(__name__)
 padding_size = 128 if bool(int(os.getenv("MOE_PADDING", "0"))) else 0
@@ -27,6 +27,19 @@ enable_moe_align_block_size_triton = bool(
     int(os.getenv("ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON", "0"))
 )
 
+_is_cuda = torch.cuda.is_available() and torch.version.cuda
+_is_rocm = torch.cuda.is_available() and torch.version.hip
+
+if _is_cuda:
+    from sgl_kernel import gelu_and_mul, silu_and_mul
+
+    from sglang.srt.layers.quantization.fp8_kernel import (
+        sglang_per_token_group_quant_fp8,
+    )
+
+if _is_cuda or _is_rocm:
+    from sgl_kernel import moe_align_block_size as sgl_moe_align_block_size
+
 
 @triton.jit
 def fused_moe_kernel(
@@ -479,7 +492,10 @@ def invoke_fused_moe_kernel(
         else:
             assert len(block_shape) == 2
             block_n, block_k = block_shape[0], block_shape[1]
-            A, A_scale = per_token_group_quant_fp8(A, block_k)
+            if _is_cuda:
+                A, A_scale = sglang_per_token_group_quant_fp8(A, block_k)
+            else:
+                A, A_scale = per_token_group_quant_fp8(A, block_k)
             assert triton.cdiv(A.shape[-1], block_k) == A_scale.shape[-1]
             assert triton.cdiv(B.shape[-2], block_n) == B_scale.shape[-2]
             assert triton.cdiv(B.shape[-1], block_k) == B_scale.shape[-1]
@@ -989,9 +1005,15 @@ def fused_experts_impl(
         )
 
         if activation == "silu":
-            ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
+            if _is_cuda:
+                silu_and_mul(intermediate_cache1.view(-1, N), intermediate_cache2)
+            else:
+                ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
         elif activation == "gelu":
-            ops.gelu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
+            if _is_cuda:
+                gelu_and_mul(intermediate_cache1.view(-1, N), intermediate_cache2)
+            else:
+                ops.gelu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
         else:
             raise ValueError(f"Unsupported activation: {activation=}")
 
@@ -1079,7 +1101,7 @@ def fused_moe(
     - num_expert_group: Optional[int]: additional parameter for grouped_topk
     - topk_group: Optional[int]: additional parameter for grouped_topk
     - use_grouped_topk: If True, use grouped_topk instead of fused_topk
-        note: Deepseekv2 model uses grouped_topk
+        note: Deepseek V2/V3/R1 series models use grouped_topk
     - use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner
         products for w1 and w2. Defaults to False.
     - use_int8_w8a16 (bool): If True, use fp8 arithmetic to compute the inner

sglang/srt/layers/moe/fused_moe_triton/layer.py

@@ -298,7 +298,9 @@ class FusedMoE(torch.nn.Module):
             layer=self,
             num_experts=num_experts,
             hidden_size=hidden_size,
+            # FIXME: figure out which intermediate_size to use
             intermediate_size=self.intermediate_size_per_partition,
+            intermediate_size_per_partition=self.intermediate_size_per_partition,
             params_dtype=params_dtype,
             weight_loader=self.weight_loader,
         )

sglang/srt/layers/moe/topk.py

@@ -75,7 +75,7 @@ def fused_topk(
     return topk_weights, topk_ids
 
 
-# This is used by the Deepseek-V2 model
+# This is used by the Deepseek V2/V3/R1 series models
 @torch.compile(dynamic=True, backend=get_compiler_backend())
 def grouped_topk(
     hidden_states: torch.Tensor,

sglang/srt/layers/quantization/__init__.py

@@ -1,10 +1,13 @@
 # Adapted from https://raw.githubusercontent.com/vllm-project/vllm/v0.5.5/vllm/model_executor/layers/quantization/__init__.py
+from typing import Callable, Dict, Optional, Type
 
-from typing import Dict, Type
-
+import torch
 from vllm.model_executor.layers.quantization.aqlm import AQLMConfig
 from vllm.model_executor.layers.quantization.awq import AWQConfig
-from vllm.model_executor.layers.quantization.awq_marlin import AWQMarlinConfig
+from vllm.model_executor.layers.quantization.awq_marlin import (
+    AWQMarlinConfig,
+    AWQMoEMethod,
+)
 from vllm.model_executor.layers.quantization.bitsandbytes import BitsAndBytesConfig
 from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors import (
     CompressedTensorsConfig,
@@ -73,21 +76,61 @@ def gptq_get_quant_method(self, layer, prefix):
 
 
 def awq_get_quant_method(self, layer, prefix):
+    from vllm.model_executor.layers.quantization.awq import is_layer_skipped_awq
     from vllm.model_executor.layers.quantization.awq_marlin import (
         AWQMarlinLinearMethod,
         AWQMoEMethod,
     )
 
-    from sglang.srt.layers.linear import LinearBase
+    from sglang.srt.layers.linear import LinearBase, UnquantizedLinearMethod
     from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
+    from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
 
-    if isinstance(layer, LinearBase):
+    if isinstance(layer, LinearBase) or (
+        isinstance(layer, ParallelLMHead) and self.lm_head_quantized
+    ):
+        if is_layer_skipped_awq(prefix, self.modules_to_not_convert):
+            return UnquantizedLinearMethod()
         return AWQMarlinLinearMethod(self)
     elif isinstance(layer, FusedMoE):
         return AWQMoEMethod(self)
     return None
 
 
+original_awq_moe_method_apply = AWQMoEMethod.apply
+
+
+def awq_moe_method_apply(
+    self,
+    layer: torch.nn.Module,
+    x: torch.Tensor,
+    router_logits: torch.Tensor,
+    top_k: int,
+    renormalize: bool,
+    use_grouped_topk: bool = False,
+    topk_group: Optional[int] = None,
+    num_expert_group: Optional[int] = None,
+    custom_routing_function: Optional[Callable] = None,
+    scoring_func: str = "softmax",
+    e_score_correction_bias: Optional[torch.Tensor] = None,
+    **kwargs,
+):
+    return original_awq_moe_method_apply(
+        self,
+        layer,
+        x,
+        router_logits,
+        top_k,
+        renormalize,
+        use_grouped_topk,
+        topk_group,
+        num_expert_group,
+        custom_routing_function,
+        scoring_func,
+        e_score_correction_bias,
+    )
+
+
 def patch_vllm_linear_base_isinstance():
     import builtins
 
@@ -107,8 +150,11 @@ def patch_vllm_linear_base_isinstance():
 
 def apply_monkey_patches():
     """Apply all monkey patches in one place."""
+    from vllm.model_executor.layers.quantization.awq_marlin import AWQMoEMethod
+
     setattr(GPTQMarlinConfig, "get_quant_method", gptq_get_quant_method)
     setattr(AWQMarlinConfig, "get_quant_method", awq_get_quant_method)
+    setattr(AWQMoEMethod, "apply", awq_moe_method_apply)
 
 
 patch_vllm_linear_base_isinstance()

sglang/srt/layers/quantization/configs/N=1536,K=1536,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

sglang/srt/layers/quantization/configs/N=1536,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,164 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    }
+}