sglang 0.5.2rc0__py3-none-any.whl → 0.5.2rc1__py3-none-any.whl

sglang/srt/mem_cache/storage/hf3fs/storage_hf3fs.py

@@ -113,6 +113,8 @@ def synchronized():
 
 
 class HiCacheHF3FS(HiCacheStorage):
+    """HiCache backend that stores KV cache pages in HF3FS files."""
+
     default_env_var: str = "SGLANG_HICACHE_HF3FS_CONFIG_PATH"
 
     def __init__(
@@ -176,15 +178,32 @@ class HiCacheHF3FS(HiCacheStorage):
         dtype: torch.dtype,
         storage_config: HiCacheStorageConfig = None,
     ) -> "HiCacheHF3FS":
+        """Create a HiCacheHF3FS instance from environment configuration.
+
+        Environment:
+            - Uses env var stored in `HiCacheHF3FS.default_env_var` to locate a JSON config.
+            - Falls back to a local single-machine config when the env var is not set.
+
+        Raises:
+            ValueError: If MLA Model is requested without global metadata server or required keys are missing.
+        """
         from sglang.srt.mem_cache.storage.hf3fs.mini_3fs_metadata_server import (
             Hf3fsGlobalMetadataClient,
             Hf3fsLocalMetadataClient,
         )
 
-        rank = storage_config.tp_rank if storage_config is not None else 0
+        if storage_config is not None:
+            rank, is_mla_model = storage_config.tp_rank, storage_config.is_mla_model
+        else:
+            rank, is_mla_model = 0, False
+
+        mla_unsupported_msg = f"MLA model is not supported without global metadata server, please refer to https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/mem_cache/storage/hf3fs/docs/deploy_sglang_3fs_multinode.md"
 
         config_path = os.getenv(HiCacheHF3FS.default_env_var)
         if not config_path:
+            if is_mla_model:
+                raise ValueError(mla_unsupported_msg)
+
             return HiCacheHF3FS(
                 rank=rank,
                 file_path=f"/data/hicache.{rank}.bin",
@@ -214,25 +233,27 @@ class HiCacheHF3FS(HiCacheStorage):
             raise ValueError(f"Missing required keys in config: {missing_keys}")
 
         # Choose metadata client based on configuration
-        is_mla_model = False
-        if "metadata_server_url" in config and config["metadata_server_url"]:
+        if config.get("metadata_server_url"):
             # Use global metadata client to connect to metadata server
             metadata_server_url = config["metadata_server_url"]
             metadata_client = Hf3fsGlobalMetadataClient(metadata_server_url)
 
-            # Enable MLA optimization only when using the global metadata client
-            is_mla_model = storage_config.is_mla_model if storage_config else False
             logger.info(
                 f"Using global metadata client with server url: {metadata_server_url}"
             )
         else:
+            # Enable MLA optimization only when using the global metadata client
+            if is_mla_model:
+                raise ValueError(mla_unsupported_msg)
+
             # Use local metadata client for single-machine deployment
             metadata_client = Hf3fsLocalMetadataClient()
 
+        rank_for_path = 0 if is_mla_model else rank
         return HiCacheHF3FS(
             rank=rank,
             # Let all ranks use the same file path for MLA model
-            file_path=f"{config['file_path_prefix']}.{rank if not is_mla_model else 0}.bin",
+            file_path=f"{config['file_path_prefix']}.{rank_for_path}.bin",
             file_size=int(config["file_size"]),
             numjobs=int(config["numjobs"]),
             bytes_per_page=bytes_per_page,

sglang/srt/models/deepseek_v2.py

@@ -2185,6 +2185,8 @@ class DeepseekV2ForCausalLM(nn.Module):
             disable_reason = "Only Deepseek V3/R1 on NV-platform with capability >= 80 can use shared experts fusion optimization."
         elif get_moe_expert_parallel_world_size() > 1:
             disable_reason = "Deepseek V3/R1 can not use shared experts fusion optimization under expert parallelism."
+        elif self.quant_config.get_name() == "w4afp8":
+            disable_reason = "Deepseek V3/R1 W4AFP8 model uses different quant method for routed experts and shared experts."
 
         if disable_reason is not None:
             global_server_args_dict["disable_shared_experts_fusion"] = True
@@ -2496,6 +2498,9 @@ class DeepseekV2ForCausalLM(nn.Module):
             ckpt_up_proj_name="up_proj",
             num_experts=self.config.n_routed_experts + self.num_fused_shared_experts,
         )
+        # Params for special naming rules in mixed-precision models, for example:
+        # model.layers.xx.mlp.experts.xx.w1.input_scale. For details,
+        # see https://huggingface.co/Barrrrry/DeepSeek-R1-W4AFP8/blob/main.
         if self.quant_config and self.quant_config.get_name() == "w4afp8":
             expert_params_mapping += FusedMoE.make_expert_input_scale_params_mapping(
                 num_experts=self.config.n_routed_experts

sglang/srt/models/gpt_oss.py

@@ -193,8 +193,9 @@ class GptOssSparseMoeBlock(nn.Module):
         return ans
 
 
-def _enable_fused_set_kv_buffer():
-    return _is_cuda
+def _enable_fused_set_kv_buffer(forward_batch: ForwardBatch):
+    """Enable fused set_kv_buffer only on CUDA with bfloat16 KV cache."""
+    return _is_cuda and forward_batch.token_to_kv_pool.dtype == torch.bfloat16
 
 
 # TODO maybe move to a model-common utils
@@ -341,7 +342,7 @@ class GptOssAttention(nn.Module):
                     layer=self.attn,
                     forward_batch=forward_batch,
                 )
-                if _enable_fused_set_kv_buffer()
+                if _enable_fused_set_kv_buffer(forward_batch)
                 else None
             ),
         )
@@ -355,7 +356,7 @@ class GptOssAttention(nn.Module):
         attn_output = self.attn(
             *inner_state,
             sinks=self.sinks,
-            save_kv_cache=not _enable_fused_set_kv_buffer(),
+            save_kv_cache=not _enable_fused_set_kv_buffer(forward_batch),
         )
         output, _ = self.o_proj(attn_output)
         return output

sglang/srt/models/longcat_flash.py

@@ -651,9 +651,6 @@ class LongcatFlashForCausalLM(nn.Module):
                         ).T
                 else:
                     w = self_attn.kv_b_proj.weight
-                # NOTE(HandH1998): Since `bmm_fp8` only supports per-tensor scale, we have to requantize `self_attn.kv_b_proj`.
-                # This may affect the accuracy of fp8 model.
-                # Fix deepseek v3 blockwise bmm by using deep_gemm
                 use_deep_gemm_bmm = False
 
                 if w.dtype in (
@@ -790,6 +787,9 @@ class LongcatFlashForCausalLM(nn.Module):
                         self.config.hidden_size / self.config.kv_lora_rank
                     ) ** 0.5
 
+        # TODO(linguoyuan) EPMoE not support DEEPGEMM_BLACKWELL, DeepEP needs to be supported in the future
+        deep_gemm_wrapper.DEEPGEMM_SCALE_UE8M0 = False
+
         if (
             deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM
             and deep_gemm_wrapper.DEEPGEMM_SCALE_UE8M0
@@ -804,24 +804,35 @@ class LongcatFlashForCausalLM(nn.Module):
         for layer_id in range(self.config.num_hidden_layers):
             layer = self.model.layers[layer_id]
             for i in range(2):
-                for module in [
-                    layer.self_attn[i].fused_qkv_a_proj_with_mqa,
-                    layer.self_attn[i].q_b_proj,
-                    layer.self_attn[i].kv_b_proj,
-                    layer.self_attn[i].o_proj,
-                ]:
-                    requant_weight_ue8m0_inplace(
-                        module.weight, module.weight_scale_inv, weight_block_size
-                    )
+                self_attn = layer.self_attn[i]
+                module_list = [
+                    self_attn.kv_b_proj,
+                    self_attn.o_proj,
+                ]
+
+                if self.config.q_lora_rank is not None:
+                    module_list.append(self_attn.fused_qkv_a_proj_with_mqa)
+                    module_list.append(self_attn.q_b_proj)
+                else:
+                    module_list.append(self_attn.kv_a_proj_with_mqa)
+                    module_list.append(self_attn.q_proj)
+
+                for module in module_list:
+                    if hasattr(module, "weight_scale_inv"):
+                        requant_weight_ue8m0_inplace(
+                            module.weight, module.weight_scale_inv, weight_block_size
+                        )
+
                 mlp = layer.mlps[i]
                 assert isinstance(mlp, LongcatFlashMLP)
                 for module in [
                     mlp.gate_up_proj,
                     mlp.down_proj,
                 ]:
-                    requant_weight_ue8m0_inplace(
-                        module.weight, module.weight_scale_inv, weight_block_size
-                    )
+                    if hasattr(module, "weight_scale_inv"):
+                        requant_weight_ue8m0_inplace(
+                            module.weight, module.weight_scale_inv, weight_block_size
+                        )
 
         for layer_id in range(self.config.num_hidden_layers):
             experts = layer.mlp.experts

sglang/srt/models/longcat_flash_nextn.py

@@ -344,9 +344,6 @@ class LongcatFlashForCausalLMNextN(LongcatFlashForCausalLM):
                 ).T
         else:
             w = self_attn.kv_b_proj.weight
-        # NOTE(HandH1998): Since `bmm_fp8` only supports per-tensor scale, we have to requantize `self_attn.kv_b_proj`.
-        # This may affect the accuracy of fp8 model.
-        # Fix deepseek v3 blockwise bmm by using deep_gemm
         use_deep_gemm_bmm = False
         if w.dtype in (
             torch.float8_e4m3fn,
@@ -480,24 +477,35 @@ class LongcatFlashForCausalLMNextN(LongcatFlashForCausalLM):
     def _weight_requant_ue8m0(self):
         weight_block_size = self.quant_config.weight_block_size
         layer = self.model.decoder
-        for module in [
-            layer.self_attn.fused_qkv_a_proj_with_mqa,
-            layer.self_attn.q_b_proj,
-            layer.self_attn.kv_b_proj,
-            layer.self_attn.o_proj,
-        ]:
-            requant_weight_ue8m0_inplace(
-                module.weight, module.weight_scale_inv, weight_block_size
-            )
+        self_attn = layer.self_attn
+        module_list = [
+            self_attn.kv_b_proj,
+            self_attn.o_proj,
+        ]
+
+        if self.config.q_lora_rank is not None:
+            module_list.append(self_attn.fused_qkv_a_proj_with_mqa)
+            module_list.append(self_attn.q_b_proj)
+        else:
+            module_list.append(self_attn.kv_a_proj_with_mqa)
+            module_list.append(self_attn.q_proj)
+
+        for module in module_list:
+            if hasattr(module, "weight_scale_inv"):
+                requant_weight_ue8m0_inplace(
+                    module.weight, module.weight_scale_inv, weight_block_size
+                )
+
         mlp = layer.mlps
         assert isinstance(mlp, LongcatFlashMLP)
         for module in [
             mlp.gate_up_proj,
             mlp.down_proj,
         ]:
-            requant_weight_ue8m0_inplace(
-                module.weight, module.weight_scale_inv, weight_block_size
-            )
+            if hasattr(module, "weight_scale_inv"):
+                requant_weight_ue8m0_inplace(
+                    module.weight, module.weight_scale_inv, weight_block_size
+                )
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [

sglang/srt/utils.py

@@ -2787,16 +2787,6 @@ def lru_cache_frozenset(maxsize=128):
     return decorator
 
 
-def get_worker_ids_from_req_rids(rids):
-    if isinstance(rids, list):
-        worker_ids = [int(rid.split("_")[0]) for rid in rids]
-    elif isinstance(rids, str):
-        worker_ids = [int(rids.split("_")[0])]
-    else:
-        worker_ids = []
-    return worker_ids
-
-
 def get_origin_rid(rid):
     return rid.split("_", 1)[1] if "_" in rid else rid
 

sglang/test/test_cutlass_w4a8_moe.py

@@ -1,6 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
 
-from typing import Optional
+from typing import Literal, Optional
 
 import pytest
 import torch
@@ -25,7 +25,7 @@ def pack_int4_values_to_int8(int4_values_interleaved: torch.Tensor) -> torch.Ten
     return packed_tensor.to(torch.int8)
 
 
-def pack_interleave(num_experts, ref_weight, ref_scale):
+def pack_interleave(num_experts, ref_weight, ref_scale, alignment=4):
     n, k = ref_weight.shape[1], ref_weight.shape[2]
 
     weight = pack_int4_values_to_int8(ref_weight.cpu()).cuda()
@@ -33,11 +33,16 @@ def pack_interleave(num_experts, ref_weight, ref_scale):
     w_q = w_q.contiguous()
 
     scale_interleaved = ref_scale.reshape(
-        ref_scale.shape[0], ref_scale.shape[1], (ref_scale.shape[2] // 4), 4
+        ref_scale.shape[0],
+        ref_scale.shape[1],
+        (ref_scale.shape[2] // alignment),
+        alignment,
     )  # [E, N, K/4, 4]
     scale_interleaved = scale_interleaved.permute(0, 2, 1, 3)  # [E, K/4, N, 4]
     scale_interleaved = scale_interleaved.reshape(
-        ref_scale.shape[0], ref_scale.shape[2] // 4, ref_scale.shape[1] * 4
+        ref_scale.shape[0],
+        ref_scale.shape[2] // alignment,
+        ref_scale.shape[1] * alignment,
     )  # [E, K/4, N*4]
     w_scale = scale_interleaved.contiguous()
 
@@ -48,12 +53,17 @@ def pack_interleave(num_experts, ref_weight, ref_scale):
 @pytest.mark.parametrize("N", [2048])
 @pytest.mark.parametrize("K", [7168])
 @pytest.mark.parametrize("E", [256])
-@pytest.mark.parametrize("ep_size", [8])
+@pytest.mark.parametrize("tp_size", [8])
+@pytest.mark.parametrize("use_ep_moe", [True, False])
 @pytest.mark.parametrize("topk", [8])
 @pytest.mark.parametrize("group_size", [128])
 @pytest.mark.parametrize("dtype", [torch.bfloat16])
-def test_cutlass_w4a8_moe(M, N, K, E, ep_size, topk, group_size, dtype):
-    local_e = E // ep_size
+def test_cutlass_w4a8_moe(M, N, K, E, tp_size, use_ep_moe, topk, group_size, dtype):
+    if use_ep_moe:
+        local_e = E // tp_size
+    else:  # tp mode
+        local_e = E
+        N = N // tp_size
 
     debug = False
     if debug:
@@ -87,7 +97,10 @@ def test_cutlass_w4a8_moe(M, N, K, E, ep_size, topk, group_size, dtype):
         )
 
     w1_q, w1_scale = pack_interleave(local_e, ref_weight_1, scale_1)
-    w2_q, w2_scale = pack_interleave(local_e, ref_weight_2, scale_2)
+    if use_ep_moe:
+        w2_q, w2_scale = pack_interleave(local_e, ref_weight_2, scale_2)
+    else:
+        w2_q, w2_scale = pack_interleave(local_e, ref_weight_2, scale_2, 1)
 
     device = "cuda"
     a_strides1 = torch.full((local_e, 3), K, device=device, dtype=torch.int64)
@@ -265,7 +278,9 @@ def ref(
 
         gate, fc1 = fc1.chunk(2, dim=-1)
         fc1 = fc1 * torch.nn.functional.silu(gate)
-        act = (fc1 / pre_quant_scale_2.float()).to(torch.float8_e4m3fn)
+        act = torch.clamp((fc1 / pre_quant_scale_2.float()), -448.0, 448.0).to(
+            torch.float8_e4m3fn
+        )
         act = act.to(dtype)
 
         w2 = ref_weight_2[e_idx]

sglang/version.py

@@ -1 +1 @@
-__version__ = "0.5.2rc0"
+__version__ = "0.5.2rc1"

{sglang-0.5.2rc0.dist-info → sglang-0.5.2rc1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sglang
-Version: 0.5.2rc0
+Version: 0.5.2rc1
 Summary: SGLang is yet another fast serving framework for large language models and vision language models.
 License:                                  Apache License
                                    Version 2.0, January 2004
@@ -257,7 +257,7 @@ Requires-Dist: uvloop; extra == "runtime-common"
 Requires-Dist: xgrammar==0.1.23; extra == "runtime-common"
 Provides-Extra: srt
 Requires-Dist: sglang[runtime_common]; extra == "srt"
-Requires-Dist: sgl-kernel==0.3.7.post1; extra == "srt"
+Requires-Dist: sgl-kernel==0.3.8; extra == "srt"
 Requires-Dist: torch==2.8.0; extra == "srt"
 Requires-Dist: torchaudio==2.8.0; extra == "srt"
 Requires-Dist: torchvision; extra == "srt"

{sglang-0.5.2rc0.dist-info → sglang-0.5.2rc1.dist-info}/RECORD

@@ -9,7 +9,7 @@ sglang/global_config.py,sha256=ZMTux_PsGnvkyJ0kTFwhTdbnFwIjjpGDogut_9Lu4Vo,1732
 sglang/launch_server.py,sha256=mDXfwha8LHpWQJekcCosR98QhCQsbmilsBlI5jAIgg0,420
 sglang/profiler.py,sha256=JCpZzlDhahoiPlPi5IG3n7GFrQHxfHEB6ELie3Ck55w,4397
 sglang/utils.py,sha256=dC2PNkKYTgDHsNrWdZJ74GvaXGSHCeIk_aZ-TA89OhY,16380
-sglang/version.py,sha256=wVQ3e58PHT2hhCLFcOLsyR4-dgdre7yd49m4mJK4CO8,25
+sglang/version.py,sha256=Yk9OPhzYNwVtFzj5deHSdkGtcxF0FXLnKJ9OFFikW9M,25
 sglang/eval/llama3_eval.py,sha256=mLNRZJIqV4CfqrY8UGnJEcHw2Xsyr1eyYZgFSUFYr1g,9997
 sglang/eval/loogle_eval.py,sha256=-CC2s2kh5qUoDrHRkQVkC_jNvBgNojXbf456ny5s78s,4557
 sglang/lang/api.py,sha256=rcp3GeoyZhmJ0GDLPRkuZNcxd0TBJy_wfUDpcmQoqW8,7210
@@ -46,7 +46,7 @@ sglang/srt/reasoning_parser.py,sha256=HEWAeFzPA_Jn3a44BYCz61QNV6kAvX46Y0tR8csAUg
 sglang/srt/server_args.py,sha256=qEh8ykOglDMHh3GvyUhG0oOSJq_tH8vUYtDzSukoOtk,104043
 sglang/srt/torch_memory_saver_adapter.py,sha256=K_eTx0UU84MHSTXI3iqYLdHV4IWtJMJ2FKdGFJR8v1E,2417
 sglang/srt/two_batch_overlap.py,sha256=UykF5nC2rja3Hvmu0D9glqKdVRIEhQGPV84Jm7veopQ,34150
-sglang/srt/utils.py,sha256=K0GxDnxVKs5LeqbCiphtGpGd5w4kPzzQsqgHkT1e0Ic,95522
+sglang/srt/utils.py,sha256=3qKij1k6uj0Ch-gErdsfXJFUiCU941L0ePoqnIqxZvk,95252
 sglang/srt/warmup.py,sha256=zldxhMlXpclRAJXmfBjJNUJd1eDizVdysibBvQyTVuA,1782
 sglang/srt/configs/__init__.py,sha256=3GdmJ2DUiNq1zNs3yOILwZzL0J8fK-h8k2P5YVgxEI0,833
 sglang/srt/configs/chatglm.py,sha256=j-b0YkdYUmQm2y1kNmMJtKeACxWKmBbvNNkDWbs6kbI,2907
@@ -60,7 +60,7 @@ sglang/srt/configs/kimi_vl.py,sha256=4W7VQI3pr888ZsFA2SqCQo4mI0seXTOrGQ-x3oTvWew
 sglang/srt/configs/kimi_vl_moonvit.py,sha256=hx2Rt4JSFbvy2HUTeLjBpge87m8M6ITAhqsgdNf_Jd4,1163
 sglang/srt/configs/load_config.py,sha256=qs-AxuplouBx2tsv9KGBOLZPbwzuVA4vbktbGP_cRp8,3309
 sglang/srt/configs/longcat_flash.py,sha256=Qp25xJVLq2K72Z80cXhcJxtqhagAdiPySDoevuT0Sno,3589
-sglang/srt/configs/model_config.py,sha256=n28KH8pqOqQNyyOwvYEXpkDZ1z8tzPe1sfV4dGmPrHc,31607
+sglang/srt/configs/model_config.py,sha256=0oEbC4bxtfPckBuY_p6uXHB1vDXxhkvJCpr9cPbExwI,31712
 sglang/srt/configs/step3_vl.py,sha256=_Otgnym57DVgB_kZ__8c1_Ys5gSalA_K0ZuVjcG51T0,4845
 sglang/srt/configs/update_config.py,sha256=GEf-XhL8JPrbX9-Hz8V7S3M6YTg76DVdIhc_4YdMDtc,6291
 sglang/srt/configs/utils.py,sha256=3nHUfisMs_Ltuhv8OZTNCJp63YJKJVF43h1QZB1zqx8,670
@@ -108,7 +108,7 @@ sglang/srt/disaggregation/nixl/conn.py,sha256=eSof87fG21Dd4COszfnbeXIxne3TWvw0mS
 sglang/srt/distributed/__init__.py,sha256=jFOcyt-wFAPMBUAf9zkZalNQlt-4rqmT6pCKBz1E4qo,149
 sglang/srt/distributed/communication_op.py,sha256=IBnFUdMftK_VSTMMMitGveonorFUUVNL4guqO31cMSc,1130
 sglang/srt/distributed/naive_distributed.py,sha256=5Kcfapzz61G3TtScTZrHoWa4bf6Vr27GlMcBAGMz7tQ,3260
-sglang/srt/distributed/parallel_state.py,sha256=RnondGDsfsbZxi2o8H1UivaoOMqDHiKVtDVb9HuOW44,65704
+sglang/srt/distributed/parallel_state.py,sha256=NsWEw341ew7dElC9BQ3vBLzaLVTDKCmCkKIl37b72dg,65717
 sglang/srt/distributed/utils.py,sha256=aaCxATncLGnVgB0WlGpBdee0behKW8Dy_dakqcuKSaQ,8497
 sglang/srt/distributed/device_communicators/cuda_wrapper.py,sha256=3jvPG-Ow5UBLiXhfx8T8snR7crSZbPpARAggsDPWq7k,7038
 sglang/srt/distributed/device_communicators/custom_all_reduce.py,sha256=Q1kkKPKFPV0QMmKLyjOBlOnX8-Pr4UeGBZYkG6j0gc0,16570
@@ -124,7 +124,7 @@ sglang/srt/distributed/device_communicators/shm_broadcast.py,sha256=IrSrnpZnii0E
 sglang/srt/distributed/device_communicators/xpu_communicator.py,sha256=ajW6132BvA6jkeipEIgN27TFycI0U06Ih2Z8WNjlA4s,1593
 sglang/srt/entrypoints/EngineBase.py,sha256=yKN76witT2jz1zhmLHmPNLGMpK2UiOTaKQ2KPD8l99U,2594
 sglang/srt/entrypoints/context.py,sha256=aD-94xkD0komuGO5gtYUoJKCHdc4hAipMxQt04yVRGA,8030
-sglang/srt/entrypoints/engine.py,sha256=7CBZly2Z1ekrBNbGHS187Yb0pj-Uuqw813s7oLrci8E,33496
+sglang/srt/entrypoints/engine.py,sha256=wNEYxQTVFHt9EvMzQr5zutX9Cb7RDDn64c8Xckuwhsg,33490
 sglang/srt/entrypoints/harmony_utils.py,sha256=01T-A5GBUm2b306PcxNEg2rfx4cykBcqNYrzcXTWBlc,13590
 sglang/srt/entrypoints/http_server.py,sha256=_GEk6RgxlMWYUNXOx9he2OIFOs1-Qan1NrSm0EAGJ3M,49649
 sglang/srt/entrypoints/http_server_engine.py,sha256=_--j4U04OeJLlnnv1f0XmCd_Ry0z1FlhkrbePX8rYV0,4938
@@ -220,7 +220,7 @@ sglang/srt/layers/attention/wave_ops/prefill_attention.py,sha256=viTUit0rxjVV5Ua
 sglang/srt/layers/moe/__init__.py,sha256=63TxUpSiUpVg1SDY1zdlTg3WFJzAc7WSndOViOmUv4E,835
 sglang/srt/layers/moe/cutlass_moe.py,sha256=JKJED-4709ndP5AwhQ7Vi04GJjw5d9Xl_mWOsZPZ3U4,14298
 sglang/srt/layers/moe/cutlass_moe_params.py,sha256=9NRCmgP_Ug3gGqCcpi-x-QRbLjCNpw8792gKXwZsbEU,6522
-sglang/srt/layers/moe/cutlass_w4a8_moe.py,sha256=rWyHMg0kYZhIdLWTJeC7pqRG9ywfeKqi47-OekNeF4Y,7306
+sglang/srt/layers/moe/cutlass_w4a8_moe.py,sha256=pUvYkbm3kD5IBjVuJeehU_hvEvdNcYcO2eRZXDc6iLI,7005
 sglang/srt/layers/moe/fused_moe_native.py,sha256=8SAToE4B-22H5JsENZgJ1Io6QfE5-D9ItWLtbGksFQQ,3372
 sglang/srt/layers/moe/rocm_moe_utils.py,sha256=07Z99bTV3B-b2Cbm-odhGpx2twxtnVpYvaDMBE1K3LM,4555
 sglang/srt/layers/moe/router.py,sha256=eUNu_Uz5VB2FOZzZyYuZo5pokCVBS17_fcjHQbmvDSE,12181
@@ -228,10 +228,13 @@ sglang/srt/layers/moe/topk.py,sha256=r8pE6eJ8dprfZxaw8VcfrRvSp6_xDNPkr0tajmG_CZc
 sglang/srt/layers/moe/utils.py,sha256=omh9E6sF-KtrELEsE9y01Ash7FfQlstyE-97UtpH8qk,5998
 sglang/srt/layers/moe/ep_moe/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 sglang/srt/layers/moe/ep_moe/kernels.py,sha256=Acg6dW-zVDQ95vuBVuzxb5SUFFxhLCewk_tVSZeuma8,46158
-sglang/srt/layers/moe/ep_moe/layer.py,sha256=izQx9iy51r8nTla8_yfEnfQUzaW07RdIskXh2rJWAf8,27969
-sglang/srt/layers/moe/fused_moe_triton/__init__.py,sha256=4NXZHbCw-G-uSnNUj4up0yh3xBDPnT-x0pdoIr0lku8,831
-sglang/srt/layers/moe/fused_moe_triton/fused_moe.py,sha256=JKp89Cdl8a85tKzmJa4ah5VtxZzRNBSbY4cnsAtnvio,60803
-sglang/srt/layers/moe/fused_moe_triton/layer.py,sha256=jub7moptP9UjyRYJ2gPvY8HEnE41d_oFbZgjVvNsJxQ,39409
+sglang/srt/layers/moe/ep_moe/layer.py,sha256=uAHuiAILb3XQi9t1JGMElfzvsG9cKIDMXs6ezEGhGvY,27776
+sglang/srt/layers/moe/fused_moe_triton/__init__.py,sha256=tis0ZJmih7gKHCurbLtY_o-bY3K4MOzQLYLC3ftIOf0,977
+sglang/srt/layers/moe/fused_moe_triton/fused_moe.py,sha256=q_L-eZh9pBP3KZLipv9cRe2WpezNrHMz8NzMy97TV0Q,23178
+sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_config.py,sha256=fFnRcXA6r0gnD_7EQmb8NacQJRJ8YHsmMZw2khlClTQ,7687
+sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_kernels.py,sha256=z10ZJLrFFhgB5_cEFrvMt6bnwe46T5Vzz4nvMl4ErGU,27776
+sglang/srt/layers/moe/fused_moe_triton/layer.py,sha256=TI3p8FOJJjpoyE7VMngs9BPeiiN9M9XHZJRWVTWtiH0,39637
+sglang/srt/layers/moe/fused_moe_triton/moe_align_block_size.py,sha256=U93mxPPU2RP7d3QRSvoG3OCtUQIK8YhQfHf1ZGeREmY,3284
 sglang/srt/layers/moe/fused_moe_triton/triton_kernels_moe.py,sha256=Ai06BZ7uxMnk0nPWQelgvi1rV9Z72FetRo6p7E3rsYs,10986
 "sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_1_0/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json",sha256=iNGsE2ZeVnQEnN4A8UJ9Jv0d3hbRF2MJ9oBgjup5Szk,2737
 "sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_1_0/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json",sha256=JJN0hryyLr5Zv3dSS7C8cPFhAwTT6XxUVnBGMZvV6JA,2752
@@ -455,7 +458,7 @@ sglang/srt/layers/quantization/petit_utils.py,sha256=-gy4zMhqNoIA1R0n7-5C0efV54j
 sglang/srt/layers/quantization/qoq.py,sha256=jM96uJfKz3vo1B6SSIrQXNT3vnD3UJrepK2eY-tSQU4,8139
 sglang/srt/layers/quantization/unquant.py,sha256=CxsOEYqYIfanKHo0ooFArbVz2ueX3vBWnYHE8gtnzdQ,12996
 sglang/srt/layers/quantization/utils.py,sha256=d4eaS4-Z4q3GRgb4HDMr2EoWEdQCefrVdJufK6n_NQY,18509
-sglang/srt/layers/quantization/w4afp8.py,sha256=345w-gCDes3dGqCUAmjyzXZemsIEWc_swJ7FyJyzFTI,11392
+sglang/srt/layers/quantization/w4afp8.py,sha256=Cm_KjJQu-XKqSKh6usJ5tNXARm6F5n_C2P1c1NuYdC0,11720
 sglang/srt/layers/quantization/w8a8_fp8.py,sha256=wzJi5jeTnbf-01iehOVTTCu_262rlMh9AQ2rogKWBmo,9981
 sglang/srt/layers/quantization/w8a8_int8.py,sha256=cOKbhWxFlv6hw5dSs8ExPBXFcTwudwR26m3CAoOhCSs,35342
 sglang/srt/layers/quantization/compressed_tensors/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -619,7 +622,7 @@ sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a
 "sglang/srt/layers/quantization/configs/N=7168,K=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128, 128].json",sha256=FImA-TJ_tQDjqwoNWxS--sRDoKDXf9gamlME3tkxH58,3252
 "sglang/srt/layers/quantization/configs/N=7168,K=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json",sha256=FFBjSWlpKXMxfAUUYUqXbOK_Hd7qBeBsfbcaa9uB4qY,3249
 sglang/srt/layers/quantization/deep_gemm_wrapper/__init__.py,sha256=esJMd0Yuj68t6QYOpmIFuiWP2J2dxTMC4bRBNH0Xk6I,26
-sglang/srt/layers/quantization/deep_gemm_wrapper/compile_utils.py,sha256=8BxWqX1WT7SvRRccBmx8CBVGiL_FIa923OlDYwzDcDM,7907
+sglang/srt/layers/quantization/deep_gemm_wrapper/compile_utils.py,sha256=PI9r3PHCMK9EgpoFuBgR4jvokA5sBz8zyC47ps3wet4,8164
 sglang/srt/layers/quantization/deep_gemm_wrapper/configurer.py,sha256=8C4xyO58N3Zl8h-fYddUpXDs3mnSqjYbaLt12cPB2XA,778
 sglang/srt/layers/quantization/deep_gemm_wrapper/entrypoint.py,sha256=vCj5vdAshEB9mAgSUYXhgJ0bd1Ithmu_n4-m_IWUbd4,2531
 sglang/srt/layers/quantization/quark/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -646,14 +649,14 @@ sglang/srt/lora/triton_ops/sgemm_lora_b.py,sha256=VqCAFvUtq_l-0RGIkx3W_fzD55QcW2
 sglang/srt/managers/cache_controller.py,sha256=WEaM01U0al13oSU7AHvwEBcmOy-1SmtPWbikgrPAx6g,33130
 sglang/srt/managers/configure_logging.py,sha256=8sNXZ2z9pBWOwn-X3wyz013Ob8Nbm1zDxRkxoZjH-l4,1633
 sglang/srt/managers/data_parallel_controller.py,sha256=VvEkidmYCGAFTv6upEHf_tyTvkM8xLhCyOnLJGWL6BQ,14592
-sglang/srt/managers/detokenizer_manager.py,sha256=YTl8ytxNf3an5sOAUjKTGcEKBJMKF7kQiViAw6eOoMw,13213
+sglang/srt/managers/detokenizer_manager.py,sha256=mVEzkaCTu0MUe8QuwQC9uReOJgTWVQYXcItlDS2gYK4,11594
 sglang/srt/managers/io_struct.py,sha256=ebAKRqBYj4slW25-dWs51JAW6caPcvuIMlSoJJZd8wo,40503
 sglang/srt/managers/mm_utils.py,sha256=J5hahCJ7HAKhJtGzV6PRa5HB9B2NjYwWGYMqIgiGC60,29444
-sglang/srt/managers/multi_tokenizer_mixin.py,sha256=smdzSwH-EW4zDNTJtkl6nXq6O_xH5EfW3iU3iSScUoM,22851
+sglang/srt/managers/multi_tokenizer_mixin.py,sha256=5HbUePDkWeEua_500sEJdKYp9CfRITKUAeFEBp8OR7k,24706
 sglang/srt/managers/multimodal_processor.py,sha256=cnWpu2G79v1a6FJB_FriLxESgGUbfC3GptLmeRVVgew,1801
 sglang/srt/managers/schedule_batch.py,sha256=9lhBnf-siQjsThD8FRyhiF50N-LOE7dfoPr1uzklnjQ,77170
 sglang/srt/managers/schedule_policy.py,sha256=vh9BQW9tBv80LW4JApLE6smU2m4gy6cAEI25HVXiS60,22383
-sglang/srt/managers/scheduler.py,sha256=7euNU5d089V4uCN7gDXHEz0eUzewQ5V7LrwAlvP5tKc,111061
+sglang/srt/managers/scheduler.py,sha256=Ipdj1_nDJDUNWb9Hx2W4NNndvhwRkn6B9rBGmJhDklc,111243
 sglang/srt/managers/scheduler_input_blocker.py,sha256=zP8xU_UmU2H0AB6sEqvivDcDB1QDgTauNDYDIJ7Nez4,3683
 sglang/srt/managers/scheduler_metrics_mixin.py,sha256=31BbukSyUUPMNHLqRlS5sEiAv7Gi5VHFQ1TgxvcAdcw,10054
 sglang/srt/managers/scheduler_output_processor_mixin.py,sha256=CLwF58GTV7PkyYUoC-R7ROA-cZt8di9_9n2pajacxUY,31511
@@ -672,10 +675,10 @@ sglang/srt/mem_cache/base_prefix_cache.py,sha256=hLS2ncTMAz7Kpdk5pNwn5c6g8b61_K9
 sglang/srt/mem_cache/chunk_cache.py,sha256=jbJeEEZ5_WYEF_AnDZIAu2sMD4hAGAd_24F980fjVwU,3199
 sglang/srt/mem_cache/flush_cache.py,sha256=GYcxmNXh4hsMpFfNOuCTpKilW7guZwTtAg_usVeM3J0,979
 sglang/srt/mem_cache/hicache_storage.py,sha256=eidtVRRsT8efqaEECXRLcxwJ7OomstzOctbqTucOLNg,7307
-sglang/srt/mem_cache/hiradix_cache.py,sha256=aL05axnc7qDy6FWMLbj76GaYIYjqooMEFYbSKiypJMA,28527
+sglang/srt/mem_cache/hiradix_cache.py,sha256=k7P_MHQnFfvFj41T3WseTQQR0Hpun4VXrqtqnjKZHtI,29155
 sglang/srt/mem_cache/lora_radix_cache.py,sha256=4NbK0Rc15kTiCOSsthyM6WsZarMK2MNhtsNqYMakJ5k,14369
 sglang/srt/mem_cache/memory_pool.py,sha256=EGZtQsgXDOuySeEBdV6jBIt9ZDvqMsErNPe6P5ifHh8,39540
-sglang/srt/mem_cache/memory_pool_host.py,sha256=d8TJqmOILaeub4maP09y20MwYf7KKobzv7pVrZ18CiA,26119
+sglang/srt/mem_cache/memory_pool_host.py,sha256=Uy2sGFEVupj7ZBWRFBgnkn0s5Yb8NzmKre9IbBQ-ut8,26189
 sglang/srt/mem_cache/multimodal_cache.py,sha256=zPnQLQhBZ6zsUpCQPSoNkrB9EEvpoDQS4mU7c3sRWjE,2171
 sglang/srt/mem_cache/radix_cache.py,sha256=U0KEjQNcwhdFWnXYTMhDkxpiazAD1ttbADEkGPHiiTU,19116
 sglang/srt/mem_cache/radix_cache_cpp.py,sha256=YuYt4xNzfmNNSkwaAk3VLWcZI74_XQobIyEgt96f7Qs,9284
@@ -683,8 +686,8 @@ sglang/srt/mem_cache/swa_radix_cache.py,sha256=hV5OOmAqu4UT80BVsu7lhGewcbq-u__JK
 sglang/srt/mem_cache/cpp_radix_tree/radix_tree.py,sha256=tQZpz-H6HxNAQe9mYqt6aX8mfDZyu_fbJTcCNK-Ns1M,7179
 sglang/srt/mem_cache/storage/hf3fs/client_hf3fs.py,sha256=WV5M9UPPLUMMjU3Mpp6HBqyCQpGBuMyj-N5PwQcNo5k,5178
 sglang/srt/mem_cache/storage/hf3fs/hf3fs_utils.cpp,sha256=dDYPtxmL6OFhEyv_GuTgzkTDpBjcD0EK8kphnn7eaGc,1133
-sglang/srt/mem_cache/storage/hf3fs/mini_3fs_metadata_server.py,sha256=OfQMFwbbQg0pbC7rJktg5wz3YFhNsNOjsXCsAjrh6iU,16738
-sglang/srt/mem_cache/storage/hf3fs/storage_hf3fs.py,sha256=RlVAloClfk_Cw4TjfokHl75YdgmtJ1FT7KdGA2QLXDk,13990
+sglang/srt/mem_cache/storage/hf3fs/mini_3fs_metadata_server.py,sha256=lJ9Jp2kOgQQaUihxlO9l-W8fYfOMa1h1y0WAtzzFGLM,17850
+sglang/srt/mem_cache/storage/hf3fs/storage_hf3fs.py,sha256=2sa0fgXRB0rkOqGbewRtO11A8ES0DhlKZi7nZmD330I,14873
 sglang/srt/mem_cache/storage/hf3fs/test_hf3fs_utils.py,sha256=g2h0Woa2CK39pRTigHCVR_hq095dTaC70lCIAIs9THM,1089
 sglang/srt/mem_cache/storage/mooncake_store/mooncake_store.py,sha256=S0UAG86sVMDc4-LvzSZrLKeW0fec5D7py_sqe3JtXEk,10000
 sglang/srt/mem_cache/storage/mooncake_store/unit_test.py,sha256=jMdJgU41VscKS-mn_AVZE8LQbb_3MWQ57i9V3OCWLI4,1007
@@ -712,7 +715,7 @@ sglang/srt/models/dbrx.py,sha256=117gwrB_o8VlACwDH5YlXkPfuS7t2Mh-nyJvpNpIYxs,159
 sglang/srt/models/deepseek.py,sha256=M7nyHcAbTIs8vc1g6u00oy22-pp704Xp4T3clBzu4xM,17460
 sglang/srt/models/deepseek_janus_pro.py,sha256=td8xGs6ARfJ8AQCYwUhMOZoWigrAs7m3trF5-kXCqik,70418
 sglang/srt/models/deepseek_nextn.py,sha256=uoiJxAVl9DX7bQcjyuouAZz6W-cB1HMubcirTBghyR4,6115
-sglang/srt/models/deepseek_v2.py,sha256=s5xrkzRtLujCKs-LUbvGl91c_QNZ4jO5Z0_6ER4tHdc,107544
+sglang/srt/models/deepseek_v2.py,sha256=IOSLn1PV3GUsNmP10nd661EOlDI3ThGZBjHS7fSEnIE,107953
 sglang/srt/models/deepseek_vl2.py,sha256=7X3gI4DE7guUtNJvMLf06jO8MUHKa2Aan8evZg5hsXc,13061
 sglang/srt/models/ernie4.py,sha256=7dP7_d0i86tv8Mp21Ew9RdsFr6TskIkm8UtUd8zFo00,16067
 sglang/srt/models/ernie4_eagle.py,sha256=E7f-ygCcY8pIoyTMKFyVC2uRcG59-_RaQTO4gd5_uWI,7223
@@ -732,7 +735,7 @@ sglang/srt/models/glm4v.py,sha256=WtDvH4691H6IKrcGBrkBpweSm_iFK61bMEfvqpbblmc,23
 sglang/srt/models/glm4v_moe.py,sha256=qgJtWyRI0dy9qcTMXHxhBZp-ghfghV3PLhpOXgSAW60,17313
 sglang/srt/models/gpt2.py,sha256=kclhxEs8oJk1KCyhmAqo7rZqecVGGHYkc-a1WZi3aIk,9841
 sglang/srt/models/gpt_bigcode.py,sha256=1D6bi8Zu760gCRZkvdLHFcg8kCkY35ARwQYaMDtYhl4,10307
-sglang/srt/models/gpt_oss.py,sha256=NoqSVjOlTzPR-TRpfViAAGyrk3lMIgyjL4RYVQQlhRY,44283
+sglang/srt/models/gpt_oss.py,sha256=6g7PE6EiOLDipmhV-XKmBcuoCGbhH6SU6NwsSoB_FJM,44469
 sglang/srt/models/granite.py,sha256=8q92shxVPAp_cJDohJATffSGd7Z0Oi-vF5jpY6DlK4s,19840
 sglang/srt/models/granitemoe.py,sha256=j1rgZ62CbBioECjUblDCw_NneDQgY_QJODsI0fqXVO8,13779
 sglang/srt/models/grok.py,sha256=8KCR13LtdnhswJrIwgb0sdFW0OxCA8GQ0fbN8gb856E,40518
@@ -753,8 +756,8 @@ sglang/srt/models/llama_embedding.py,sha256=zq-_lNu35VBFc7eemiam0zdkGIE8fzrgk5OW
 sglang/srt/models/llama_reward.py,sha256=LF2nqMV5XOrljGjAwJg43mBv3z6Q040I2EYlgZeCp8k,4681
 sglang/srt/models/llava.py,sha256=xzYip_BAwpzSIdZre43LZiyTpFISa0ZCLdO6LUSbaCg,37702
 sglang/srt/models/llavavid.py,sha256=-CSk0RJ2MQeb81sh-RISeVJFaI-XWY6nR6_I594MkME,12818
-sglang/srt/models/longcat_flash.py,sha256=ZtVmyxmEG_ckDnWEj-rwa2q99sgB59XHMqiSpnmcKdw,40705
-sglang/srt/models/longcat_flash_nextn.py,sha256=HxK_W0kcTrJu2WXMOJNYHajq8s-8qbHCWo7YBT-9e6E,29489
+sglang/srt/models/longcat_flash.py,sha256=gyElibQ9q_qq4tBA7eah0f3bLSzvMEpgD_n6CcRyG9M,41055
+sglang/srt/models/longcat_flash_nextn.py,sha256=oVy776-AE5z43C472oqHlcl9NfzDDEaQgGB1msZNdgc,29644
 sglang/srt/models/mimo.py,sha256=Mp-iFp4YHuiuq-H8enUF5K5QbMnVcvEa6mURH6vM3yM,6140
 sglang/srt/models/mimo_mtp.py,sha256=jSmqJAXu7G3OO7jW1oa2suI4H_Yl7u5ZT7w4lHFbHhg,7292
 sglang/srt/models/minicpm.py,sha256=CzBJyZtfMpp8jvlEl29DHI7HLVq-CxuqP1UHwxIbaUI,14567
@@ -857,7 +860,7 @@ sglang/test/test_block_fp8_deep_gemm_blackwell.py,sha256=Hnhq4kkyINHb4ONedkp5Kf7
 sglang/test/test_block_fp8_ep.py,sha256=n4X6ZKwuUUbV5Ofjg64ptlaFGI1LbRXDfFiJW1ELHgY,10546
 sglang/test/test_custom_ops.py,sha256=PenQ8zM1wj5xwiVEPVzD37pO-x90aOfFMpCRZenaKsY,5709
 sglang/test/test_cutlass_moe.py,sha256=ax-IYPm5tZjkZft0q8swHnzerOI4LB4JOLFaKGJVE-k,9629
-sglang/test/test_cutlass_w4a8_moe.py,sha256=E0ffQOR84bEFz6KFNbczH2LRhQqkeC9HPp76u0vzdwo,8809
+sglang/test/test_cutlass_w4a8_moe.py,sha256=Ku9VCXEMJ3BwXtvb1A3FzY-zek-S-A4thWyX6m-1v-A,9219
 sglang/test/test_deepep_utils.py,sha256=749ysTBGNzh6rYUCJhhZBtZpeD15eWTeNHYCytcvZtc,7448
 sglang/test/test_dynamic_grad_mode.py,sha256=L76yUCuk_ymNpXD2CmO8r2GiGjIvD_gtTsuFDs2NolI,1638
 sglang/test/test_fp4_moe.py,sha256=rJLkKW3glBMvI5Ed0LltOHi8zCReMa-WB50p5zGm6J4,10189
@@ -871,8 +874,8 @@ sglang/test/attention/test_flashattn_backend.py,sha256=_rTG849FwQdVTyGKkqhczaOqn
 sglang/test/attention/test_flashattn_mla_backend.py,sha256=g4O50WblTpM7_Gq2b76k0i25_z01BOUBQ4i6PmyxpO4,10774
 sglang/test/attention/test_prefix_chunk_info.py,sha256=hpoDe2wfSa6RlUbfyri_c0iyBTb35UXGL9I2Xh6jamM,7772
 sglang/test/attention/test_trtllm_mla_backend.py,sha256=quZ6SYuEH7J1YMcF8YO1_bwSNMz1gecpWRGauYjbUeA,42055
-sglang-0.5.2rc0.dist-info/licenses/LICENSE,sha256=FJXh51fvTQklojUFY89XVLsjxRcBqOxPs8XNy-2uZ0c,11346
-sglang-0.5.2rc0.dist-info/METADATA,sha256=6S-9JaZ2NMphKGxZVRyfEQWUwyjRY3I0a3HY4qNLU00,28157
-sglang-0.5.2rc0.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-sglang-0.5.2rc0.dist-info/top_level.txt,sha256=yxhh3pYQkcnA7v3Bg889C2jZhvtJdEincysO7PEB09M,7
-sglang-0.5.2rc0.dist-info/RECORD,,
+sglang-0.5.2rc1.dist-info/licenses/LICENSE,sha256=FJXh51fvTQklojUFY89XVLsjxRcBqOxPs8XNy-2uZ0c,11346
+sglang-0.5.2rc1.dist-info/METADATA,sha256=mMYLEpjZJzf_6puNSLlKkwaPpNRJ9nJ7mAsmRgSvb3k,28151
+sglang-0.5.2rc1.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+sglang-0.5.2rc1.dist-info/top_level.txt,sha256=yxhh3pYQkcnA7v3Bg889C2jZhvtJdEincysO7PEB09M,7
+sglang-0.5.2rc1.dist-info/RECORD,,