PyPI - sglang - Versions diffs - 0.4.5.post3__py3-none-any.whl → 0.4.6.post1__py3-none-any.whl - Mend

sglang 0.4.5.post3py3-none-any.whl → 0.4.6.post1py3-none-any.whl

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (97) hide show

sglang/bench_one_batch.py +19 -3
sglang/bench_serving.py +8 -9
sglang/compile_deep_gemm.py +45 -4
sglang/srt/code_completion_parser.py +1 -1
sglang/srt/configs/deepseekvl2.py +1 -1
sglang/srt/configs/model_config.py +9 -3
sglang/srt/constrained/llguidance_backend.py +78 -61
sglang/srt/conversation.py +34 -1
sglang/srt/disaggregation/decode.py +67 -13
sglang/srt/disaggregation/fake/__init__.py +1 -0
sglang/srt/disaggregation/fake/conn.py +88 -0
sglang/srt/disaggregation/mini_lb.py +45 -8
sglang/srt/disaggregation/mooncake/conn.py +198 -31
sglang/srt/disaggregation/prefill.py +36 -12
sglang/srt/disaggregation/utils.py +16 -2
sglang/srt/entrypoints/engine.py +9 -0
sglang/srt/entrypoints/http_server.py +35 -4
sglang/srt/function_call_parser.py +77 -5
sglang/srt/layers/attention/base_attn_backend.py +3 -0
sglang/srt/layers/attention/cutlass_mla_backend.py +278 -0
sglang/srt/layers/attention/flashattention_backend.py +28 -10
sglang/srt/layers/attention/flashmla_backend.py +8 -11
sglang/srt/layers/attention/utils.py +1 -1
sglang/srt/layers/attention/vision.py +2 -0
sglang/srt/layers/layernorm.py +38 -16
sglang/srt/layers/logits_processor.py +2 -2
sglang/srt/layers/moe/fused_moe_native.py +2 -4
sglang/srt/layers/moe/fused_moe_triton/configs/E=128,N=192,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
sglang/srt/layers/moe/fused_moe_triton/configs/E=128,N=192,device_name=NVIDIA_H20.json +146 -0
sglang/srt/layers/moe/fused_moe_triton/configs/E=128,N=192,device_name=NVIDIA_H200.json +146 -0
sglang/srt/layers/moe/fused_moe_triton/configs/E=128,N=384,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
sglang/srt/layers/moe/fused_moe_triton/configs/E=128,N=384,device_name=NVIDIA_H20.json +146 -0
sglang/srt/layers/moe/fused_moe_triton/configs/E=128,N=384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
sglang/srt/layers/moe/fused_moe_triton/configs/E=128,N=384,device_name=NVIDIA_H200.json +146 -0
sglang/srt/layers/moe/fused_moe_triton/configs/E=128,N=768,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
sglang/srt/layers/moe/fused_moe_triton/configs/E=128,N=768,device_name=NVIDIA_H20.json +146 -0
sglang/srt/layers/moe/fused_moe_triton/configs/E=128,N=768,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
sglang/srt/layers/moe/fused_moe_triton/configs/E=128,N=768,device_name=NVIDIA_H200.json +146 -0
sglang/srt/layers/moe/fused_moe_triton/configs/E=128,N=96,device_name=NVIDIA_H20.json +146 -0
sglang/srt/layers/moe/fused_moe_triton/configs/E=264,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +41 -41
sglang/srt/layers/moe/fused_moe_triton/configs/E=272,N=128,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
sglang/srt/layers/moe/fused_moe_triton/fused_moe.py +20 -17
sglang/srt/layers/moe/fused_moe_triton/layer.py +15 -17
sglang/srt/layers/pooler.py +6 -0
sglang/srt/layers/quantization/awq.py +5 -1
sglang/srt/layers/quantization/deep_gemm.py +17 -10
sglang/srt/layers/quantization/fp8.py +20 -22
sglang/srt/layers/quantization/fp8_utils.py +2 -2
sglang/srt/layers/quantization/int8_kernel.py +32 -1
sglang/srt/layers/radix_attention.py +13 -3
sglang/srt/layers/rotary_embedding.py +170 -126
sglang/srt/managers/data_parallel_controller.py +10 -3
sglang/srt/managers/io_struct.py +7 -0
sglang/srt/managers/mm_utils.py +85 -28
sglang/srt/managers/multimodal_processors/base_processor.py +14 -1
sglang/srt/managers/multimodal_processors/deepseek_vl_v2.py +9 -2
sglang/srt/managers/multimodal_processors/gemma3.py +2 -5
sglang/srt/managers/multimodal_processors/janus_pro.py +2 -2
sglang/srt/managers/multimodal_processors/minicpm.py +4 -3
sglang/srt/managers/multimodal_processors/qwen_vl.py +38 -13
sglang/srt/managers/schedule_batch.py +38 -12
sglang/srt/managers/scheduler.py +41 -28
sglang/srt/managers/scheduler_output_processor_mixin.py +25 -9
sglang/srt/managers/tokenizer_manager.py +5 -1
sglang/srt/managers/tp_worker.py +3 -3
sglang/srt/managers/tp_worker_overlap_thread.py +9 -4
sglang/srt/mem_cache/memory_pool.py +87 -0
sglang/srt/model_executor/cuda_graph_runner.py +4 -3
sglang/srt/model_executor/forward_batch_info.py +51 -95
sglang/srt/model_executor/model_runner.py +19 -25
sglang/srt/models/deepseek.py +12 -2
sglang/srt/models/deepseek_nextn.py +101 -6
sglang/srt/models/deepseek_v2.py +144 -70
sglang/srt/models/deepseek_vl2.py +9 -4
sglang/srt/models/gemma3_causal.py +1 -1
sglang/srt/models/llama4.py +0 -1
sglang/srt/models/minicpmo.py +5 -1
sglang/srt/models/mllama4.py +2 -2
sglang/srt/models/qwen2_5_vl.py +3 -6
sglang/srt/models/qwen2_vl.py +3 -7
sglang/srt/models/roberta.py +178 -0
sglang/srt/openai_api/adapter.py +50 -11
sglang/srt/openai_api/protocol.py +2 -0
sglang/srt/reasoning_parser.py +25 -1
sglang/srt/server_args.py +31 -24
sglang/srt/speculative/eagle_draft_cuda_graph_runner.py +3 -3
sglang/srt/torch_memory_saver_adapter.py +10 -1
sglang/srt/utils.py +5 -1
sglang/test/runners.py +6 -13
sglang/test/send_one.py +84 -28
sglang/test/test_utils.py +74 -18
sglang/version.py +1 -1
{sglang-0.4.5.post3.dist-info → sglang-0.4.6.post1.dist-info}/METADATA +5 -6
{sglang-0.4.5.post3.dist-info → sglang-0.4.6.post1.dist-info}/RECORD +97 -80
{sglang-0.4.5.post3.dist-info → sglang-0.4.6.post1.dist-info}/WHEEL +1 -1
{sglang-0.4.5.post3.dist-info → sglang-0.4.6.post1.dist-info}/licenses/LICENSE +0 -0
{sglang-0.4.5.post3.dist-info → sglang-0.4.6.post1.dist-info}/top_level.txt +0 -0

sglang/srt/layers/layernorm.py CHANGED Viewed

@@ -22,8 +22,6 @@ import torch.nn as nn
 from sglang.srt.custom_op import CustomOp
 from sglang.srt.utils import is_cuda, is_hip
-logger = logging.getLogger(__name__)
 _is_cuda = is_cuda()
 _is_hip = is_hip()
@@ -36,19 +34,9 @@ if _is_cuda:
     )
 if _is_hip:
+    from vllm._custom_ops import fused_add_rms_norm, rms_norm
-    from aiter.ops.rmsnorm import rms_norm, rmsnorm2d_fwd_with_add
-    rmsnorm = rms_norm
-    def fused_add_rmsnorm(
-        x: torch.Tensor,
-        residual: torch.Tensor,
-        w: torch.Tensor,
-        eps: float,
-    ) -> Tuple[torch.Tensor, torch.Tensor]:
-        rmsnorm2d_fwd_with_add(x, x, residual, residual, w, eps)
-        return x, residual
+logger = logging.getLogger(__name__)
 class RMSNorm(CustomOp):
@@ -61,23 +49,49 @@ class RMSNorm(CustomOp):
         self.weight = nn.Parameter(torch.ones(hidden_size))
         self.variance_epsilon = eps
+    def forward(self, *args, **kwargs):
+        if torch.compiler.is_compiling():
+            return self.forward_native(*args, **kwargs)
+        if _is_cuda:
+            return self.forward_cuda(*args, **kwargs)
+        elif _is_hip:
+            return self.forward_hip(*args, **kwargs)
+        else:
+            return self.forward_native(*args, **kwargs)
     def forward_cuda(
         self,
         x: torch.Tensor,
         residual: Optional[torch.Tensor] = None,
     ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
         if residual is not None:
             fused_add_rmsnorm(x, residual, self.weight.data, self.variance_epsilon)
             return x, residual
         out = rmsnorm(x, self.weight.data, self.variance_epsilon)
         return out
+    def forward_hip(
+        self,
+        x: torch.Tensor,
+        residual: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        if not x.is_contiguous():
+            # NOTE: Romove this if aiter kernel supports discontinuous input
+            x = x.contiguous()
+        if residual is not None:
+            fused_add_rms_norm(x, residual, self.weight.data, self.variance_epsilon)
+            return x, residual
+        out = torch.empty_like(x)
+        rms_norm(out, x, self.weight.data, self.variance_epsilon)
+        return out
     def forward_native(
         self,
         x: torch.Tensor,
         residual: Optional[torch.Tensor] = None,
     ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        if not x.is_contiguous():
+            x = x.contiguous()
         orig_dtype = x.dtype
         x = x.to(torch.float32)
         if residual is not None:
@@ -103,6 +117,14 @@ class GemmaRMSNorm(CustomOp):
         self.weight = nn.Parameter(torch.zeros(hidden_size))
         self.variance_epsilon = eps
+    def forward(self, *args, **kwargs):
+        if torch.compiler.is_compiling():
+            return self.forward_native(*args, **kwargs)
+        if _is_cuda:
+            return self.forward_cuda(*args, **kwargs)
+        else:
+            return self.forward_native(*args, **kwargs)
     def forward_native(
         self,
         x: torch.Tensor,
@@ -156,6 +178,6 @@ class Gemma3RMSNorm(nn.Module):
 if not (_is_cuda or _is_hip):
     logger.info(
-        "sgl-kernel is not available on Non-NV platforms. Fallback to other kernel libraries."
+        "sgl-kernel layernorm implementation is not available on current platform. Fallback to other kernel libraries."
     )
     from vllm.model_executor.layers.layernorm import GemmaRMSNorm, RMSNorm

sglang/srt/layers/logits_processor.py CHANGED Viewed

@@ -335,13 +335,13 @@ class LogitsProcessor(nn.Module):
                     aux_pruned_states = torch.cat(aux_pruned_states, dim=-1)
                     hidden_states_to_store = (
                         aux_pruned_states[sample_indices]
-                        if sample_indices
+                        if sample_indices is not None
                         else aux_pruned_states
                     )
                 else:
                     hidden_states_to_store = (
                         pruned_states[sample_indices]
-                        if sample_indices
+                        if sample_indices is not None
                         else pruned_states
                     )
             else:

sglang/srt/layers/moe/fused_moe_native.py CHANGED Viewed

@@ -8,6 +8,7 @@ from typing import Callable, Optional
 import torch
 from torch.nn import functional as F
+from sglang.srt.layers.activation import GeluAndMul, SiluAndMul
 from sglang.srt.layers.moe.topk import select_experts
@@ -30,7 +31,7 @@ def fused_moe_forward_native(
 ) -> torch.Tensor:
     if apply_router_weight_on_input:
-        raise NotImplementedError
+        raise NotImplementedError()
     topk_weights, topk_ids = select_experts(
         hidden_states=x,
@@ -75,9 +76,6 @@ def moe_forward_native(
     activation: str = "silu",
     routed_scaling_factor: Optional[float] = None,
 ) -> torch.Tensor:
-    from sglang.srt.layers.activation import GeluAndMul, SiluAndMul
     topk_weights, topk_ids = select_experts(
         hidden_states=x,
         router_logits=router_logits,

sglang/srt/layers/moe/fused_moe_triton/configs/E=128,N=192,device_name=NVIDIA_H100_80GB_HBM3.json ADDED Viewed

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

sglang/srt/layers/moe/fused_moe_triton/configs/E=128,N=192,device_name=NVIDIA_H20.json ADDED Viewed

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

sglang/srt/layers/moe/fused_moe_triton/configs/E=128,N=192,device_name=NVIDIA_H200.json ADDED Viewed

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

sglang 0.4.5.post3__py3-none-any.whl → 0.4.6.post1__py3-none-any.whl

sglang 0.4.5.post3py3-none-any.whl → 0.4.6.post1py3-none-any.whl