sglang 0.5.1.post3__py3-none-any.whl → 0.5.2__py3-none-any.whl

sglang/srt/layers/moe/cutlass_w4a8_moe.py

@@ -91,18 +91,10 @@ def cutlass_w4a8_moe(
     assert w1_q.shape[0] == w2_q.shape[0], "Expert number mismatch"
     assert w1_q.shape[0] == w1_scale.shape[0], "w1 scales expert number mismatch"
     assert w1_q.shape[0] == w2_scale.shape[0], "w2 scales expert number mismatch"
-    assert (
-        w1_scale.shape[1] == w1_q.shape[2] * 2 / 512
-        and w1_scale.shape[2] == w1_q.shape[1] * 4
-    ), "W1 scale shape mismatch"
-    assert (
-        w2_scale.shape[1] == w2_q.shape[2] * 2 / 512
-        and w2_scale.shape[2] == w2_q.shape[1] * 4
-    ), "W2 scale shape mismatch"
 
     assert a_strides1.shape[0] == w1_q.shape[0], "A Strides 1 expert number mismatch"
     assert b_strides1.shape[0] == w1_q.shape[0], "B Strides 1 expert number mismatch"
-    assert a_strides2.shape[0] == w2_q.shape[0], "A Strides 2 expert number  mismatch"
+    assert a_strides2.shape[0] == w2_q.shape[0], "A Strides 2 expert number mismatch"
     assert b_strides2.shape[0] == w2_q.shape[0], "B Strides 2 expert number mismatch"
     num_experts = w1_q.size(0)
     m = a.size(0)
@@ -155,8 +147,8 @@ def cutlass_w4a8_moe(
         k,
     )
 
-    c1 = torch.empty((m * topk, n * 2), device=device, dtype=torch.half)
-    c2 = torch.zeros((m * topk, k), device=device, dtype=torch.half)
+    c1 = torch.empty((m * topk, n * 2), device=device, dtype=torch.bfloat16)
+    c2 = torch.zeros((m * topk, k), device=device, dtype=torch.bfloat16)
 
     cutlass_w4a8_moe_mm(
         c1,
@@ -174,7 +166,7 @@ def cutlass_w4a8_moe(
         topk,
     )
 
-    intermediate = torch.empty((m * topk, n), device=device, dtype=torch.half)
+    intermediate = torch.empty((m * topk, n), device=device, dtype=torch.bfloat16)
     silu_and_mul(c1, intermediate)
 
     intermediate_q = torch.empty(

sglang/srt/layers/moe/ep_moe/kernels.py

@@ -1362,3 +1362,77 @@ def moe_ep_deepgemm_preprocess(
         gateup_input,
         gateup_input_scale,
     )
+
+
+@triton.jit
+def compute_identity_kernel(
+    top_k,
+    hidden_states_ptr,
+    expert_scales_ptr,
+    num_tokens,
+    output_ptr,
+    hidden_dim,
+    scales_stride,
+    BLOCK_SIZE: tl.constexpr,
+):
+    pid = tl.program_id(0)
+
+    batch_id = pid // (hidden_dim // BLOCK_SIZE)
+    dim_offset = pid % (hidden_dim // BLOCK_SIZE) * BLOCK_SIZE
+
+    if batch_id >= num_tokens or dim_offset >= hidden_dim:
+        return
+
+    h = tl.load(
+        hidden_states_ptr
+        + batch_id * hidden_dim
+        + dim_offset
+        + tl.arange(0, BLOCK_SIZE),
+        mask=(dim_offset + tl.arange(0, BLOCK_SIZE)) < hidden_dim,
+    )
+
+    result = tl.zeros([BLOCK_SIZE], dtype=tl.float32)
+    for i in range(top_k):
+        scale = tl.load(expert_scales_ptr + batch_id * scales_stride + i)
+        result += h * scale
+
+    tl.store(
+        output_ptr + batch_id * hidden_dim + dim_offset + tl.arange(0, BLOCK_SIZE),
+        result,
+        mask=(dim_offset + tl.arange(0, BLOCK_SIZE)) < hidden_dim,
+    )
+
+
+def zero_experts_compute_triton(
+    expert_indices, expert_scales, num_experts, zero_expert_type, hidden_states
+):
+    N = expert_indices.numel()
+    top_k = expert_indices.size(-1)
+    grid = lambda meta: (triton.cdiv(N, meta["BLOCK_SIZE"]),)
+
+    if zero_expert_type == "identity":
+        zero_expert_mask = expert_indices < num_experts
+        zero_expert_scales = expert_scales.clone()
+        zero_expert_scales[zero_expert_mask] = 0.0
+
+    normal_expert_mask = expert_indices >= num_experts
+    expert_indices[normal_expert_mask] = -1
+    expert_scales[normal_expert_mask] = 0.0
+
+    output = torch.zeros_like(hidden_states).to(hidden_states.device)
+    hidden_dim = hidden_states.size(-1)
+    num_tokens = hidden_states.size(0)
+
+    grid = lambda meta: (num_tokens * (hidden_dim // meta["BLOCK_SIZE"]),)
+    compute_identity_kernel[grid](
+        top_k,
+        hidden_states,
+        zero_expert_scales,
+        num_tokens,
+        output,
+        hidden_dim,
+        zero_expert_scales.stride(0),
+        BLOCK_SIZE=256,
+    )
+
+    return output

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

@@ -35,7 +35,6 @@ from sglang.srt.utils import ceil_div, dispose_tensor, get_bool_env_var, is_hip,
 
 if TYPE_CHECKING:
     from sglang.srt.layers.moe.token_dispatcher import (
-        AscendDeepEPLLOutput,
         DeepEPLLOutput,
         DeepEPNormalOutput,
         DispatchOutput,
@@ -114,9 +113,6 @@ class EPMoE(FusedMoE):
             with_bias=with_bias,
         )
 
-        self.start_expert_id = self.moe_ep_rank * self.num_local_experts
-        self.end_expert_id = self.start_expert_id + self.num_local_experts - 1
-
         self.intermediate_size = intermediate_size
 
         if isinstance(quant_config, Fp8Config):
@@ -232,7 +228,7 @@ class EPMoE(FusedMoE):
             (
                 _cast_to_e8m0_with_rounding_up(gateup_input_scale)
                 if deep_gemm_wrapper.DEEPGEMM_SCALE_UE8M0
-                else deep_gemm_wrapper.get_col_major_tma_aligned_tensor(
+                else deep_gemm_wrapper.get_mn_major_tma_aligned_tensor(
                     gateup_input_scale
                 )
             ),
@@ -289,9 +285,7 @@ class EPMoE(FusedMoE):
             (
                 down_input_scale
                 if deep_gemm_wrapper.DEEPGEMM_SCALE_UE8M0
-                else deep_gemm_wrapper.get_col_major_tma_aligned_tensor(
-                    down_input_scale
-                )
+                else deep_gemm_wrapper.get_mn_major_tma_aligned_tensor(down_input_scale)
             ),
         )
         down_output = torch.empty(
@@ -459,7 +453,7 @@ class DeepEPMoE(EPMoE):
             # in forward_aiter, we skip token permutation and unpermutation, which have been fused inside aiter kernel
             return self.forward_aiter(dispatch_output)
         if _is_npu:
-            assert DispatchOutputChecker.format_is_ascent_ll(dispatch_output)
+            assert DispatchOutputChecker.format_is_deepep(dispatch_output)
             return self.forward_npu(dispatch_output)
         if DispatchOutputChecker.format_is_deepep_normal(dispatch_output):
             assert deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM and self.use_fp8_w8a8
@@ -723,57 +717,124 @@ class DeepEPMoE(EPMoE):
 
     def forward_npu(
         self,
-        dispatch_output: DeepEPLLOutput,
+        dispatch_output: Union[DeepEPNormalOutput, DeepEPLLOutput],
     ):
-        if TYPE_CHECKING:
-            assert isinstance(dispatch_output, AscendDeepEPLLOutput)
-        hidden_states, topk_idx, topk_weights, _, seg_indptr, _ = dispatch_output
         assert self.quant_method is not None
         assert self.moe_runner_config.activation == "silu"
 
+        import torch_npu
+
+        from sglang.srt.layers.moe.token_dispatcher import DispatchOutputChecker
+
         # NOTE: Ascend's Dispatch & Combine does not support FP16
         output_dtype = torch.bfloat16
+        group_list_type = 1
 
-        pertoken_scale = hidden_states[1]
-        hidden_states = hidden_states[0]
+        def _forward_normal(dispatch_output: DeepEPNormalOutput):
+            if TYPE_CHECKING:
+                assert isinstance(dispatch_output, DeepEPNormalOutput)
+            hidden_states, _, _, num_recv_tokens_per_expert = dispatch_output
+
+            if isinstance(hidden_states, tuple):
+                per_token_scale = hidden_states[1]
+                hidden_states = hidden_states[0]
+            else:
+                # dynamic quant
+                hidden_states, per_token_scale = torch_npu.npu_dynamic_quant(
+                    hidden_states
+                )
 
-        group_list_type = 1
-        seg_indptr = seg_indptr.to(torch.int64)
+            group_list = torch.tensor(num_recv_tokens_per_expert, dtype=torch.int64).to(
+                hidden_states.device
+            )
 
-        import torch_npu
+            # gmm1: gate_up_proj
+            hidden_states = torch_npu.npu_grouped_matmul(
+                x=[hidden_states],
+                weight=[self.w13_weight],
+                scale=[self.w13_weight_scale.to(output_dtype)],
+                per_token_scale=[per_token_scale],
+                split_item=2,
+                group_list_type=group_list_type,
+                group_type=0,
+                group_list=group_list,
+                output_dtype=output_dtype,
+            )[0]
+
+            # act_fn: swiglu
+            hidden_states = torch_npu.npu_swiglu(hidden_states)
+            hidden_states, swiglu_out_scale = torch_npu.npu_dynamic_quant(hidden_states)
+
+            # gmm2: down_proj
+            hidden_states = torch_npu.npu_grouped_matmul(
+                x=[hidden_states],
+                weight=[self.w2_weight],
+                scale=[self.w2_weight_scale.to(output_dtype)],
+                per_token_scale=[swiglu_out_scale],
+                split_item=2,
+                group_list_type=group_list_type,
+                group_type=0,
+                group_list=group_list,
+                output_dtype=output_dtype,
+            )[0]
+
+            return hidden_states
 
-        # gmm1: gate_up_proj
-        hidden_states = torch_npu.npu_grouped_matmul(
-            x=[hidden_states],
-            weight=[self.w13_weight],
-            scale=[self.w13_weight_scale.to(output_dtype)],
-            per_token_scale=[pertoken_scale],
-            split_item=2,
-            group_list_type=group_list_type,
-            group_type=0,
-            group_list=seg_indptr,
-            output_dtype=output_dtype,
-        )[0]
-
-        # act_fn: swiglu
-        hidden_states = torch_npu.npu_swiglu(hidden_states)
-
-        hidden_states, swiglu_out_scale = torch_npu.npu_dynamic_quant(hidden_states)
-
-        # gmm2: down_proj
-        hidden_states = torch_npu.npu_grouped_matmul(
-            x=[hidden_states],
-            weight=[self.w2_weight],
-            scale=[self.w2_weight_scale.to(output_dtype)],
-            per_token_scale=[swiglu_out_scale],
-            split_item=2,
-            group_list_type=group_list_type,
-            group_type=0,
-            group_list=seg_indptr,
-            output_dtype=output_dtype,
-        )[0]
+        def _forward_ll(dispatch_output: DeepEPLLOutput):
+            if TYPE_CHECKING:
+                assert isinstance(dispatch_output, DeepEPLLOutput)
+            hidden_states, topk_idx, topk_weights, group_list, _ = dispatch_output
+
+            per_token_scale = hidden_states[1]
+            hidden_states = hidden_states[0]
+
+            group_list = group_list.to(torch.int64)
+
+            # gmm1: gate_up_proj
+            hidden_states = torch_npu.npu_grouped_matmul(
+                x=[hidden_states],
+                weight=[self.w13_weight],
+                split_item=2,
+                group_list_type=group_list_type,
+                group_type=0,
+                group_list=group_list,
+                output_dtype=torch.int32,
+            )[0]
+
+            # act_fn: swiglu
+            hidden_states, swiglu_out_scale = torch_npu.npu_dequant_swiglu_quant(
+                x=hidden_states,
+                weight_scale=self.w13_weight_scale.to(torch.float32),
+                activation_scale=per_token_scale,
+                bias=None,
+                quant_scale=None,
+                quant_offset=None,
+                group_index=group_list,
+                activate_left=True,
+                quant_mode=1,
+            )
 
-        return hidden_states
+            # gmm2: down_proj
+            hidden_states = torch_npu.npu_grouped_matmul(
+                x=[hidden_states],
+                weight=[self.w2_weight],
+                scale=[self.w2_weight_scale.to(output_dtype)],
+                per_token_scale=[swiglu_out_scale],
+                split_item=2,
+                group_list_type=group_list_type,
+                group_type=0,
+                group_list=group_list,
+                output_dtype=output_dtype,
+            )[0]
+
+            return hidden_states
+
+        if DispatchOutputChecker.format_is_deepep_normal(dispatch_output):
+            return _forward_normal(dispatch_output)
+        elif DispatchOutputChecker.format_is_deepep_ll(dispatch_output):
+            return _forward_ll(dispatch_output)
+        else:
+            raise ValueError(f"Not Supported DeepEP format {dispatch_output.format}")
 
 
 def get_moe_impl_class(quant_config: Optional[QuantizationConfig] = None):

sglang/srt/layers/moe/fused_moe_native.py

@@ -8,16 +8,18 @@ from torch.nn import functional as F
 
 from sglang.srt.layers.activation import GeluAndMul, SiluAndMul
 from sglang.srt.layers.moe.moe_runner import MoeRunnerConfig
+from sglang.srt.layers.moe.token_dispatcher import StandardDispatchOutput
 from sglang.srt.layers.moe.topk import StandardTopKOutput
 
 
 def fused_moe_forward_native(
     layer: torch.nn.Module,
-    x: torch.Tensor,
-    topk_output: StandardTopKOutput,
-    moe_runner_config: MoeRunnerConfig,
+    dispatch_output: StandardDispatchOutput,
 ) -> torch.Tensor:
 
+    x, topk_output = dispatch_output
+    moe_runner_config = layer.moe_runner_config
+
     if moe_runner_config.apply_router_weight_on_input:
         raise NotImplementedError()
 

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

@@ -1,16 +1,18 @@
 from contextlib import contextmanager
 from typing import Any, Dict, Optional
 
-from sglang.srt.layers.moe.fused_moe_triton.fused_moe import (
-    fused_experts,
+from sglang.srt.layers.moe.fused_moe_triton.fused_moe import fused_experts
+from sglang.srt.layers.moe.fused_moe_triton.fused_moe_triton_config import (
     get_config_file_name,
-    moe_align_block_size,
     try_get_optimal_moe_config,
 )
 from sglang.srt.layers.moe.fused_moe_triton.layer import (
     FusedMoE,
     FusedMoeWeightScaleSupported,
 )
+from sglang.srt.layers.moe.fused_moe_triton.moe_align_block_size import (
+    moe_align_block_size,
+)
 
 _config: Optional[Dict[str, Any]] = None
 

sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_3_1/E=128,N=768,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json

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

sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=129,N=352,device_name=NVIDIA_B200,dtype=fp8_w8a8.json

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