sglang 0.4.0.post2__py3-none-any.whl → 0.4.1.post1__py3-none-any.whl

sglang/test/test_block_fp8.py

@@ -0,0 +1,341 @@
+import itertools
+import unittest
+
+import torch
+
+from sglang.srt.layers.activation import SiluAndMul
+from sglang.srt.layers.moe.fused_moe_triton.fused_moe import fused_moe
+from sglang.srt.layers.quantization.fp8_kernel import (
+    per_token_group_quant_fp8,
+    w8a8_block_fp8_matmul,
+)
+
+
+# For test
+def native_per_token_group_quant_fp8(
+    x, group_size, eps=1e-10, dtype=torch.float8_e4m3fn
+):
+    """Function to perform per-token-group quantization on an input tensor `x` using native torch.
+
+    It converts the tensor values into float8 values and returns the
+    quantized tensor along with the scaling factor used for quantization.
+    Note that only `torch.float8_e4m3fn` is supported for now.
+    """
+    assert (
+        x.shape[-1] % group_size == 0
+    ), "the last dimension of `x` cannot be divisible by `group_size`"
+    assert x.is_contiguous(), "`x` is not contiguous"
+
+    finfo = torch.finfo(dtype)
+    fp8_min = finfo.min
+    fp8_max = finfo.max
+
+    x_ = x.reshape(x.numel() // group_size, group_size)
+    amax = x_.abs().max(dim=-1, keepdim=True)[0].clamp(min=eps).to(torch.float32)
+    x_s = amax / fp8_max
+    x_q = (x_ / x_s).clamp(min=fp8_min, max=fp8_max).to(dtype)
+    x_q = x_q.reshape(x.shape)
+    x_s = x_s.reshape(x.shape[:-1] + (x.shape[-1] // group_size,))
+
+    return x_q, x_s
+
+
+class TestPerTokenGroupQuantFP8(unittest.TestCase):
+    DTYPES = [torch.half, torch.bfloat16, torch.float32]
+    NUM_TOKENS = [7, 83, 2048]
+    D = [512, 4096, 5120, 13824]
+    GROUP_SIZE = [64, 128, 256, 512]
+    SEEDS = [0]
+
+    @classmethod
+    def setUpClass(cls):
+        if not torch.cuda.is_available():
+            raise unittest.SkipTest("CUDA is not available")
+        torch.set_default_device("cuda")
+
+    def _per_token_group_quant_fp8(self, num_tokens, d, dtype, group_size, seed):
+        torch.manual_seed(seed)
+
+        x = torch.rand(num_tokens, d, dtype=dtype)
+
+        with torch.inference_mode():
+            ref_out, ref_scale = native_per_token_group_quant_fp8(x, group_size)
+            out, scale = per_token_group_quant_fp8(x, group_size)
+
+        self.assertTrue(
+            torch.allclose(out.to(torch.float32), ref_out.to(torch.float32), rtol=0.15)
+        )
+        self.assertTrue(torch.allclose(scale, ref_scale))
+
+    def test_per_token_group_quant_fp8(self):
+        for params in itertools.product(
+            self.NUM_TOKENS,
+            self.D,
+            self.DTYPES,
+            self.GROUP_SIZE,
+            self.SEEDS,
+        ):
+            with self.subTest(
+                num_tokens=params[0],
+                d=params[1],
+                dtype=params[2],
+                group_size=params[3],
+                seed=params[4],
+            ):
+                self._per_token_group_quant_fp8(*params)
+
+
+# For test
+def native_w8a8_block_fp8_matmul(A, B, As, Bs, block_size, output_dtype=torch.float16):
+    """This function performs matrix multiplication with block-wise quantization using native torch.
+
+    It takes two input tensors `A` and `B` with scales `As` and `Bs`.
+    The output is returned in the specified `output_dtype`.
+    """
+
+    A = A.to(torch.float32)
+    B = B.to(torch.float32)
+    assert A.shape[-1] == B.shape[-1]
+    assert B.ndim == 2 and B.is_contiguous() and Bs.ndim == 2
+    assert len(block_size) == 2
+    block_n, block_k = block_size[0], block_size[1]
+    assert (A.shape[-1] + block_k - 1) // block_k == As.shape[-1]
+    assert A.shape[:-1] == As.shape[:-1]
+
+    M = A.numel() // A.shape[-1]
+    N, K = B.shape
+    origin_C_shape = A.shape[:-1] + (N,)
+    A = A.reshape(M, A.shape[-1])
+    As = As.reshape(M, As.shape[-1])
+    n_tiles = (N + block_n - 1) // block_n
+    k_tiles = (K + block_k - 1) // block_k
+    assert n_tiles == Bs.shape[0]
+    assert k_tiles == Bs.shape[1]
+
+    C_shape = (M, N)
+    C = torch.zeros(C_shape, dtype=torch.float32, device=A.device)
+
+    A_tiles = [A[:, i * block_k : min((i + 1) * block_k, K)] for i in range(k_tiles)]
+    B_tiles = [
+        [
+            B[
+                j * block_n : min((j + 1) * block_n, N),
+                i * block_k : min((i + 1) * block_k, K),
+            ]
+            for i in range(k_tiles)
+        ]
+        for j in range(n_tiles)
+    ]
+    C_tiles = [C[:, j * block_n : min((j + 1) * block_n, N)] for j in range(n_tiles)]
+    As_tiles = [As[:, i : i + 1] for i in range(k_tiles)]
+
+    for i in range(k_tiles):
+        for j in range(n_tiles):
+            a = A_tiles[i]
+            b = B_tiles[j][i]
+            c = C_tiles[j]
+            s = As_tiles[i] * Bs[j][i]
+            c[:, :] += torch.matmul(a, b.t()) * s
+
+    C = C.reshape(origin_C_shape).to(output_dtype)
+    return C
+
+
+class TestW8A8BlockFP8Matmul(unittest.TestCase):
+    OUT_DTYPES = [torch.float32, torch.half, torch.bfloat16]
+    M = [1, 7, 83, 512, 2048]
+    N = [128, 512, 1024, 4096, 7748, 13824]
+    K = [256, 4096, 5120, 3884, 13824]
+    # BLOCK_SIZE = [[64, 64], [64, 128], [128, 64], [128, 128]]
+    BLOCK_SIZE = [[128, 128]]
+    SEEDS = [0]
+
+    @classmethod
+    def setUpClass(cls):
+        if not torch.cuda.is_available():
+            raise unittest.SkipTest("CUDA is not available")
+        torch.set_default_device("cuda")
+
+    def _w8a8_block_fp8_matmul(self, M, N, K, block_size, out_dtype, seed):
+        torch.manual_seed(seed)
+        # NOTE(HandH1998): to avoid overflow when out_dtype = torch.half
+        factor_for_scale = 1e-2
+        fp8_info = torch.finfo(torch.float8_e4m3fn)
+        fp8_max, fp8_min = fp8_info.max, fp8_info.min
+
+        A_fp32 = (torch.rand(M, K, dtype=torch.float32) - 0.5) * 2 * fp8_max
+        A_fp8 = A_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
+
+        B_fp32 = (torch.rand(N, K, dtype=torch.float32) - 0.5) * 2 * fp8_max
+        B_fp8 = B_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
+
+        block_n, block_k = block_size[0], block_size[1]
+        n_tiles = (N + block_n - 1) // block_n
+        k_tiles = (K + block_k - 1) // block_k
+
+        As = torch.rand(M, k_tiles, dtype=torch.float32) * factor_for_scale
+        Bs = torch.rand(n_tiles, k_tiles, dtype=torch.float32) * factor_for_scale
+
+        with torch.inference_mode():
+            ref_out = native_w8a8_block_fp8_matmul(
+                A_fp8, B_fp8, As, Bs, block_size, out_dtype
+            )
+            out = w8a8_block_fp8_matmul(A_fp8, B_fp8, As, Bs, block_size, out_dtype)
+
+        self.assertTrue(
+            torch.mean(torch.abs(out.to(torch.float32) - ref_out.to(torch.float32)))
+            / torch.mean(torch.abs(ref_out.to(torch.float32)))
+            < 0.001
+        )
+
+    def test_w8a8_block_fp8_matmul(self):
+        for params in itertools.product(
+            self.M,
+            self.N,
+            self.K,
+            self.BLOCK_SIZE,
+            self.OUT_DTYPES,
+            self.SEEDS,
+        ):
+            with self.subTest(
+                M=params[0],
+                N=params[1],
+                K=params[2],
+                block_size=params[3],
+                out_dtype=params[4],
+                seed=params[5],
+            ):
+                self._w8a8_block_fp8_matmul(*params)
+
+
+# For test
+def torch_w8a8_block_fp8_moe(a, w1, w2, w1_s, w2_s, score, topk, block_shape):
+    """This function performs fused moe with block-wise quantization using native torch."""
+
+    B, D = a.shape
+    a = a.view(B, -1, D).repeat(1, topk, 1).reshape(-1, D)
+    out = torch.zeros(B * topk, w2.shape[1], dtype=a.dtype, device=a.device)
+    score = torch.softmax(score, dim=-1, dtype=torch.float32)
+    topk_weight, topk_ids = torch.topk(score, topk)
+    topk_weight = topk_weight.view(-1)
+    topk_ids = topk_ids.view(-1)
+
+    _, block_k = block_shape[0], block_shape[1]
+    a_q, a_s = native_per_token_group_quant_fp8(a, block_k)
+    # NOTE(HandH1998): Since "index_cuda" not implemented for 'Float8_e4m3fn', we need to cast `float8`` to `float32``.
+    a_q = a_q.to(torch.float32)
+    for i in range(w1.shape[0]):
+        mask = topk_ids == i
+        if mask.sum():
+            inter_out = native_w8a8_block_fp8_matmul(
+                a_q[mask], w1[i], a_s[mask], w1_s[i], block_shape, output_dtype=a.dtype
+            )
+            act_out = SiluAndMul().forward_native(inter_out)
+            act_out_q, act_out_s = native_per_token_group_quant_fp8(act_out, block_k)
+            act_out = act_out.to(torch.float32)
+            out[mask] = native_w8a8_block_fp8_matmul(
+                act_out_q, w2[i], act_out_s, w2_s[i], block_shape, output_dtype=a.dtype
+            )
+    return (
+        out.view(B, -1, w2.shape[1]) * topk_weight.view(B, -1, 1).to(out.dtype)
+    ).sum(dim=1)
+
+
+class TestW8A8BlockFP8FusedMoE(unittest.TestCase):
+    DTYPES = [torch.float32, torch.half, torch.bfloat16]
+    M = [1, 33, 64, 222, 1024 * 128]
+    N = [128, 1024, 2048]
+    K = [256, 4096, 5120]
+    E = [8, 24]
+    TOP_KS = [2, 6]
+    BLOCK_SIZE = [[64, 64], [64, 128], [128, 64], [128, 128]]
+    # BLOCK_SIZE = [[128, 128]]
+    SEEDS = [0]
+
+    @classmethod
+    def setUpClass(cls):
+        if not torch.cuda.is_available():
+            raise unittest.SkipTest("CUDA is not available")
+        torch.set_default_device("cuda")
+
+    def _w8a8_block_fp8_fused_moe(self, M, N, K, E, topk, block_size, dtype, seed):
+        torch.manual_seed(seed)
+        # NOTE(HandH1998): to avoid overflow when out_dtype = torch.half
+        factor_for_scale = 1e-2
+        fp8_info = torch.finfo(torch.float8_e4m3fn)
+        fp8_max, fp8_min = fp8_info.max, fp8_info.min
+
+        a = torch.randn((M, K), dtype=dtype) / 10
+
+        w1_fp32 = (torch.rand((E, 2 * N, K), dtype=torch.float32) - 0.5) * 2 * fp8_max
+        w1 = w1_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
+
+        w2_fp32 = (torch.rand((E, K, N), dtype=torch.float32) - 0.5) * 2 * fp8_max
+        w2 = w2_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
+
+        block_n, block_k = block_size[0], block_size[1]
+        n_tiles_w1 = (2 * N + block_n - 1) // block_n
+        n_tiles_w2 = (K + block_n - 1) // block_n
+        k_tiles_w1 = (K + block_k - 1) // block_k
+        k_tiles_w2 = (N + block_k - 1) // block_k
+
+        w1_s = (
+            torch.rand((E, n_tiles_w1, k_tiles_w1), dtype=torch.float32)
+            * factor_for_scale
+        )
+        w2_s = (
+            torch.rand((E, n_tiles_w2, k_tiles_w2), dtype=torch.float32)
+            * factor_for_scale
+        )
+
+        score = torch.randn((M, E), dtype=dtype)
+
+        with torch.inference_mode():
+            out = fused_moe(
+                a,
+                w1,
+                w2,
+                score,
+                topk,
+                renormalize=False,
+                use_fp8_w8a8=True,
+                w1_scale=w1_s,
+                w2_scale=w2_s,
+                block_shape=block_size,
+            )
+            ref_out = torch_w8a8_block_fp8_moe(
+                a, w1, w2, w1_s, w2_s, score, topk, block_size
+            )
+
+        self.assertTrue(
+            torch.mean(torch.abs(out.to(torch.float32) - ref_out.to(torch.float32)))
+            / torch.mean(torch.abs(ref_out.to(torch.float32)))
+            < 0.02
+        )
+
+    def test_w8a8_block_fp8_fused_moe(self):
+        for params in itertools.product(
+            self.M,
+            self.N,
+            self.K,
+            self.E,
+            self.TOP_KS,
+            self.BLOCK_SIZE,
+            self.DTYPES,
+            self.SEEDS,
+        ):
+            with self.subTest(
+                M=params[0],
+                N=params[1],
+                K=params[2],
+                E=params[3],
+                topk=params[4],
+                block_size=params[5],
+                dtype=params[6],
+                seed=params[7],
+            ):
+                self._w8a8_block_fp8_fused_moe(*params)
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=2)

sglang/version.py

@@ -1 +1 @@
-__version__ = "0.4.0.post2"
+__version__ = "0.4.1.post1"

{sglang-0.4.0.post2.dist-info → sglang-0.4.1.post1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: sglang
-Version: 0.4.0.post2
+Version: 0.4.1.post1
 Summary: SGLang is yet another fast serving framework for large language models and vision language models.
 License:                                  Apache License
                                    Version 2.0, January 2004
@@ -234,7 +234,6 @@ Requires-Dist: pydantic; extra == "runtime-common"
 Requires-Dist: python-multipart; extra == "runtime-common"
 Requires-Dist: pyzmq>=25.1.2; extra == "runtime-common"
 Requires-Dist: torchao>=0.7.0; extra == "runtime-common"
-Requires-Dist: gemlite; extra == "runtime-common"
 Requires-Dist: uvicorn; extra == "runtime-common"
 Requires-Dist: uvloop; extra == "runtime-common"
 Requires-Dist: xgrammar>=0.1.6; extra == "runtime-common"
@@ -244,6 +243,7 @@ Requires-Dist: torch; extra == "srt"
 Requires-Dist: vllm<=0.6.4.post1,>=0.6.3.post1; extra == "srt"
 Requires-Dist: cuda-python; extra == "srt"
 Requires-Dist: flashinfer==0.1.6; extra == "srt"
+Requires-Dist: sgl-kernel>=0.0.2.post10; extra == "srt"
 Provides-Extra: srt-hip
 Requires-Dist: sglang[runtime_common]; extra == "srt-hip"
 Requires-Dist: torch; extra == "srt-hip"
@@ -358,8 +358,8 @@ Learn more in our release blogs: [v0.2 blog](https://lmsys.org/blog/2024-07-25-s
 [Development Roadmap (2024 Q4)](https://github.com/sgl-project/sglang/issues/1487)
 
 ## Adoption and Sponsorship
-The project is supported by (alphabetically): AMD, Baseten, Etched, Hyperbolic, Jam & Tea Studios, LinkedIn, Meituan, NVIDIA, RunPod, Stanford, UC Berkeley, xAI and 01.AI.
+The project is supported by (alphabetically): AMD, Baseten, DataCrunch, Etched, Hyperbolic, Jam & Tea Studios, LinkedIn, Meituan, NVIDIA, RunPod, Stanford, UC Berkeley, xAI, 01.AI.
 
 ## Acknowledgment and Citation
 We learned from the design and reused code from the following projects: [Guidance](https://github.com/guidance-ai/guidance), [vLLM](https://github.com/vllm-project/vllm), [LightLLM](https://github.com/ModelTC/lightllm), [FlashInfer](https://github.com/flashinfer-ai/flashinfer), [Outlines](https://github.com/outlines-dev/outlines), and [LMQL](https://github.com/eth-sri/lmql).
-Please cite our paper, [SGLang: Efficient Execution of Structured Language Model Programs](https://arxiv.org/abs/2312.07104), if you find the project useful.
+Please cite the paper, [SGLang: Efficient Execution of Structured Language Model Programs](https://arxiv.org/abs/2312.07104), if you find the project useful.

{sglang-0.4.0.post2.dist-info → sglang-0.4.1.post1.dist-info}/RECORD

@@ -1,17 +1,17 @@
 sglang/__init__.py,sha256=b2oIdWzp5P8SzieeOs2TzJoN3Do3tfJbV8gZS_imVcs,1619
 sglang/api.py,sha256=NdO6cYnklnEBQBKqQjlqI8-P1EownKQ71t5ibCGhEVo,6953
 sglang/bench_latency.py,sha256=oZjSAzX7dUiSu-zdz0dkyUPo-qAX_lsXFH1gf03akgI,76
-sglang/bench_offline_throughput.py,sha256=rgMWDhA1Hai0gKBzxc0dzTWfI8l39Cyw2VOCyMt1YyY,12771
-sglang/bench_one_batch.py,sha256=aF0onHeRjy7AYVjsq1IA3rZEhUuYXuslg1fAhuvJ2yo,16120
+sglang/bench_offline_throughput.py,sha256=iQiJCK3KQDCdwU1NVbIwbtthssWzBXiIsKUDA7Z_hO0,12510
+sglang/bench_one_batch.py,sha256=jkyMhK0lqn5dRCYgAh30qZrNHP4gAbXODymBMNXK86I,15859
 sglang/bench_one_batch_server.py,sha256=-fV9FTLNNcSIy0pgYeggXedPVK0fVsXZqVQswT8OMOY,5945
-sglang/bench_serving.py,sha256=zv_EcbWno79j7WYFL2m6BfCLT6iSOfGV4uwGbDg9KQA,53141
+sglang/bench_serving.py,sha256=YQiCZreejCPBTqMmZsCB99RMi1N-Jx-dZtaafcQ8-14,53377
 sglang/check_env.py,sha256=4OqpZaEJOfBM6-vtPILto5kqDmgiZM1Koc7lK78A7CI,8427
 sglang/global_config.py,sha256=fnT0U9vlHdGaQFKN9tYTnUF4-eVW4HYQURd5zvPtrg0,1286
 sglang/launch_server.py,sha256=4y2QeSj0wVNB9MJQZeahD4ahTDU6gwqo7MPUytyFop0,403
 sglang/launch_server_llavavid.py,sha256=tGc17S1vUfLwbi1GB26oOdXxTWr7gjlqpTrPnrMRNO8,1007
 sglang/llama3_eval.py,sha256=gWSboDchIGybIce88bJlrCG0yiLZ513mw4gcutJlzGM,10017
 sglang/utils.py,sha256=23jf4Mz8E5p5a6JOkjnfYZixdjZUk88F_mZ8rZcby5Q,11597
-sglang/version.py,sha256=OUNovuQ1RrdJFYetl0e0U0556H_wiyjhVks9-l-zF94,28
+sglang/version.py,sha256=ARioq8ApVNckeQorLPVfHZeN9mlHMLbaNgLGNbGq-ys,28
 sglang/lang/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 sglang/lang/chat_template.py,sha256=cnfjjxIIcYRGRxXlJlOGnpFxFuhMHut7DS52LsOMKcA,15826
 sglang/lang/choices.py,sha256=-W1DVw9N9ZliVpvmWrzIXG4cswAah8eMQrHWzkS3D8o,6234
@@ -23,28 +23,29 @@ sglang/lang/backend/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSu
 sglang/lang/backend/anthropic.py,sha256=EXRX7xJgA5KZszX7toSLVnKzFQ5EO0Loj-YjHFtxSxg,2081
 sglang/lang/backend/base_backend.py,sha256=tdoh9YF3CyekY1BKiX9n7-aA4srDWIuA4RDJLM7q8qg,1985
 sglang/lang/backend/litellm.py,sha256=ugmL7sfUxkUHVbHtwNzHgdQAEd4UCjNQboFuE3KThcY,2450
-sglang/lang/backend/openai.py,sha256=qM7eVH_kMxnDd2rpxOH0v76KxtOJFlAwgLgWIKvFGCI,15060
+sglang/lang/backend/openai.py,sha256=ha9a2P6T80TmSgYlyIwB1qYawWkjcOgiOptkktkqa1U,15436
 sglang/lang/backend/runtime_endpoint.py,sha256=dfs-yZ1ekKmnbpZLluQHWPmMeZJKbaaZRRGYRa9eBE8,10541
 sglang/lang/backend/vertexai.py,sha256=O-iBLD-y3vq80UxnrAoJri7bxpgd-_eakZ88Cf8bEGA,4855
 sglang/srt/_custom_ops.py,sha256=Y4gyTDGhWz-W2Igq25Ojm8XFiyvkawW9I-79iwYvxJ0,3574
+sglang/srt/aio_rwlock.py,sha256=6LYtOdeTUY3hkfa1dmYkgsaF2ttrwIF3hUWz2AZ2fqw,2970
 sglang/srt/conversation.py,sha256=u9zFU8aMYzwHUbQRKU76B_T-jfLlPoxUcWG_nRbDM2I,21201
 sglang/srt/hf_transformers_utils.py,sha256=38Ms0H2-VMerOS6jnczcFtZMS6lhw9B5rSWKAfxVUfQ,7945
 sglang/srt/mm_utils.py,sha256=1ScBunw_x4W8ebM_AcJ62-1T2mfT8NlMJqdAhkF1lb0,12367
 sglang/srt/model_parallel.py,sha256=eLXZhvJ4wG6dh0FontNCIdVZvHYdWgaeY-5cu7TD9tE,6078
-sglang/srt/server.py,sha256=tEciMH_U6WIZYPUGDDM0c4BQ16cvgVdA4II-ksPZoMo,34621
+sglang/srt/server.py,sha256=vDucJl6qtEK2swzPJ_wYitaJvsI4MigMagGlBlH5V54,34033
 sglang/srt/server_args.py,sha256=LgnQ-kBJZ3E7hMMZj9bSK0mn7Bhjk1nJHxLcxl-lGTM,34572
-sglang/srt/utils.py,sha256=WWEcMJHmvlOjiqE9UicT0ZYwa2PUKDZorAk2Y8PPRBI,42039
+sglang/srt/utils.py,sha256=J8kFl6kDBwFZCM6AKaVTiqdhJKRg0JOH0pNrD1ZeWmM,41726
 sglang/srt/configs/__init__.py,sha256=_usVIXHQjft4PAJ1Y-yGQOn2QNOv501GYMlQwpGXbns,208
 sglang/srt/configs/device_config.py,sha256=dResqHjkg_dq10v6rnVpbXpvABZRB0jylOm-2_JAnx0,428
 sglang/srt/configs/exaone.py,sha256=Duxd4yQoKy8GWEzZD_kCY_OzmN_67CTJL_Kgn0eXk3g,10731
 sglang/srt/configs/load_config.py,sha256=TcPi_HY6xu5SiVZsxPOoB5pGeDUNebOk7muoUH9VBDg,3083
-sglang/srt/configs/model_config.py,sha256=OjEeigs5tMNKP-RImJk2NHVFXv-fyQfsGREWMO3rqhM,15839
+sglang/srt/configs/model_config.py,sha256=vVarlLTw9Ged1PXIwRP-R8UhiG6oaezNIZhTNuF0eQc,16070
 sglang/srt/configs/qwen2vl.py,sha256=ZjLy9v2eZY4wptUfY3CWgYKg2B5DDrkfCSyTy_Zf_bg,4351
 sglang/srt/constrained/__init__.py,sha256=UWZNVLvOT5ZBX8M36sONgDmnKtkQ0cSfhQD2jO0ATuk,786
 sglang/srt/constrained/base_grammar_backend.py,sha256=FhVm7PxhXDl0joV9NP5RjKgz7dR1dZvUAQnh0mdtvVY,2353
 sglang/srt/constrained/outlines_backend.py,sha256=CipNHNNXs8xtnJNVNe6FCwZUlSbIXbGmWVlZz3hUpFQ,6820
 sglang/srt/constrained/outlines_jump_forward.py,sha256=iZWXeR3gNYoMubLGyFmLPO4V2YsN5DiGjD71Xk9iFaE,6418
-sglang/srt/constrained/xgrammar_backend.py,sha256=4It9_GqU4UZFhxIw_7hkzpXaMPUtksk6Xfe0Agsfw7A,4620
+sglang/srt/constrained/xgrammar_backend.py,sha256=76oUFXeB29bfnEVWa1-rIrwQm5jhuMlzAX10HtAq1fQ,4887
 sglang/srt/distributed/__init__.py,sha256=__tl9Frrf3PFrSyNYcn5i-y2rL-J4-Qn6RJwrsZ4xgc,83
 sglang/srt/distributed/communication_op.py,sha256=ZoIhboZyefiAwr-1K-wF3rAFSQ4Wt-RxXpsX443Gbt4,1157
 sglang/srt/distributed/parallel_state.py,sha256=HplRH5S0AWdwSdhoHYX9_UWQZlFjh2Z1LHaz68EXlpE,47555
@@ -60,35 +61,37 @@ sglang/srt/distributed/device_communicators/shm_broadcast.py,sha256=WVxBd1QfIgRW
 sglang/srt/distributed/device_communicators/xpu_communicator.py,sha256=P3WKgddcfpUhBa-_5PvjYxH146ZE-N1cotTzEpPRKlY,1620
 sglang/srt/layers/activation.py,sha256=EboMjT9HV2tNHQ6rzpojtlkzev1lAFbhQlxMg9hwxBQ,5471
 sglang/srt/layers/custom_op_util.py,sha256=0vu-yX2wwonmO1L_o5G7SA6C-8XuhDIh9rPDvNeLhoc,922
-sglang/srt/layers/fused_moe_patch.py,sha256=DMIyrwOON7OSidKZdreL5HzMhP0AD5Ues0xdY-ADOQw,4471
 sglang/srt/layers/layernorm.py,sha256=nRQ1w1xSUcU-zlqVC61BnGG6otS5W1w9VaSzeXizrx4,4037
-sglang/srt/layers/linear.py,sha256=dF2HvqiMbhWlCjvkLFRCcgUFGhG-B0keM_CIpjvgTtg,46154
+sglang/srt/layers/linear.py,sha256=KyRFU0VcoNuN-hnQB9QQcBN9NCpeqPtLzzufIHUpV6w,47064
 sglang/srt/layers/logits_processor.py,sha256=JlOU0x8vBGIuTwHSdjR6Kly9_uzilBMv0NE_rvUx0W4,14747
 sglang/srt/layers/pooler.py,sha256=rj2lygvleBnyLCBZ8I11HGMgpfIDsT0l3PIkshJwdu4,1606
 sglang/srt/layers/radix_attention.py,sha256=E4cmvkcCdCtb6VyLNrCKy1D6VwHQ063oH3JQXPaRy6w,2178
 sglang/srt/layers/rotary_embedding.py,sha256=29tx3JNR40AoXqBa2cFGBjva9vU2xgFipETlpMaaZas,3985
 sglang/srt/layers/sampler.py,sha256=k4Op_HMkQfT7t9wgQwBVotfTUXEocrzRyQqEFnff1pc,5511
-sglang/srt/layers/torchao_utils.py,sha256=07Fe2Csdh1JiQKPGGHWkbq0-a6bV7Cq136ygdtVAhgI,3708
+sglang/srt/layers/torchao_utils.py,sha256=dQVuWNXxAvOPjr2G5BBMWqC2oKcS2B52rx-fEc_elmc,3545
 sglang/srt/layers/vocab_parallel_embedding.py,sha256=slGwLiWjuFLCUdRe-GTlfumyZpqVX9VF6No_UGOT-hA,21624
 sglang/srt/layers/attention/__init__.py,sha256=KIJhzOJWYioQE7Va4D83-V-ZUZVMZcczuNgDC3dlSRo,2583
 sglang/srt/layers/attention/double_sparsity_backend.py,sha256=RQdEKRykSLf9ilnaHmR6T7RFqh4emH_adfB3aJN2BUU,10920
-sglang/srt/layers/attention/flashinfer_backend.py,sha256=NgeigL1WiPOuOry0Gbxv-6HEcERB8Du0mBJgYcTVIAA,24943
+sglang/srt/layers/attention/flashinfer_backend.py,sha256=umD1E2zvMnPbbgvx2Ex5LQB6a4a41brjsks1M0gFMMU,26357
 sglang/srt/layers/attention/torch_native_backend.py,sha256=nQdeqWEMMH_wrod5wssDCJG-uPKm0uslvkALKqPRPQ8,10509
 sglang/srt/layers/attention/triton_backend.py,sha256=-TobyZHwlbJ5HhbFg-jgCqVOw4Y-opgEuFo-EusASQc,6264
 sglang/srt/layers/attention/triton_ops/decode_attention.py,sha256=oJ_UK1t229zF3hbTDiQe7t-X-IbM2dOxx4U2ch-vmjA,17847
 sglang/srt/layers/attention/triton_ops/double_sparsity_attention.py,sha256=1pSXfY3EEaM7iRN_uElHnAfsrJMhTFbu9fj8Z0O2PbE,21480
-sglang/srt/layers/attention/triton_ops/extend_attention.py,sha256=tZJhzqcf1KKMT8z7_32eVk_D1NHP71c-S3UNxemfAHM,11542
+sglang/srt/layers/attention/triton_ops/extend_attention.py,sha256=DWOZXSTVN5ZbcFjDjcqs-nPdUkxSwum0SVXhVKqwh2g,11688
 sglang/srt/layers/attention/triton_ops/prefill_attention.py,sha256=lojFXRZMLWkzS2Y8uxaolnQhXaWKG19mCAWaF5KQeiI,6087
-sglang/srt/layers/ep_moe/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-sglang/srt/layers/ep_moe/kernels.py,sha256=wb_S2qLxoWWgQu9coXy0XLNGvHzdZSdwXr0PGy4QySg,10940
-sglang/srt/layers/ep_moe/layer.py,sha256=uMropMhU-MaycoxSLxcfD0jZC_cuL_boRbIu86mbZjY,23034
-sglang/srt/layers/fused_moe_triton/__init__.py,sha256=PHKFqd2hPOO-g9kSMseg2g76lpg9OGXQDThWU6bt9vs,902
-sglang/srt/layers/fused_moe_triton/fused_moe.py,sha256=fLGmkY6imJYjEw9-3-jJthkMcFGMBcu9HCNIuxAzMhE,29625
-sglang/srt/layers/fused_moe_triton/layer.py,sha256=eMpbZlP3FAQxbHochis7ybZ-fsNBP0PzKF1PN0Xo7so,21517
-sglang/srt/layers/quantization/__init__.py,sha256=FgNy_zNWMWnq3lEGyCSyfLSQtcZtWlq99JilkmEDW7I,4594
+sglang/srt/layers/moe/fused_moe_native.py,sha256=8q-LFZMSCGLc2_Gltp2lH0gSb4A1WOuKQW3wo3rpj5g,1601
+sglang/srt/layers/moe/topk.py,sha256=JpeIl_-CNk0yyG3k5fmmNbbmR2_9bkKC23UoLOlMkjw,6954
+sglang/srt/layers/moe/ep_moe/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+sglang/srt/layers/moe/ep_moe/kernels.py,sha256=wb_S2qLxoWWgQu9coXy0XLNGvHzdZSdwXr0PGy4QySg,10940
+sglang/srt/layers/moe/ep_moe/layer.py,sha256=6iQU5ZjQ8IXGoQ8ZlBuJqyQxYTEem9vXI6rbVIWKlZw,22303
+sglang/srt/layers/moe/fused_moe_triton/__init__.py,sha256=h9yMFAL_bagUf-qBED8gSWdCOb7d8IdA-pE-L_nIg8E,842
+sglang/srt/layers/moe/fused_moe_triton/fused_moe.py,sha256=zXwWUtthLa9E35EvlQ9A_mnIsQyA0_NYKsUBdJqONHo,31163
+sglang/srt/layers/moe/fused_moe_triton/layer.py,sha256=BclDj5JyCy-8Bfue4broL1-IG6a4dUyggE9WQLa06sg,20575
+sglang/srt/layers/quantization/__init__.py,sha256=VPYXShHvbvkOgVBlkIqic4RhdJ1y6EZ3r34T-nZMT1k,4606
 sglang/srt/layers/quantization/base_config.py,sha256=daK9p0aijMszLUm1W4Pc33FK87MdqYK1NoWFKif-j80,4599
-sglang/srt/layers/quantization/fp8.py,sha256=3oIUPaD0PBXQyTKr44I0YJ8XXDdwyoS_-ZA97XdSxXE,24143
-sglang/srt/layers/quantization/fp8_utils.py,sha256=eJDLLDu8ZbrbE3BfFIf89JlIMPOP-14DesbeVsajW0Q,1035
+sglang/srt/layers/quantization/fp8.py,sha256=wNnpXLroIl7D98mlfCiXZPE9hrP5ricHrXY1WZBzEEo,30810
+sglang/srt/layers/quantization/fp8_kernel.py,sha256=eoO1enzD9jPC80id2oC3i8bt-LN6-4Ey223yOQ9yIPE,8792
+sglang/srt/layers/quantization/fp8_utils.py,sha256=HBJBaNcln1NrLxzw0ppUjMd6w-ryuGDDHCYJq7mRQac,4035
 sglang/srt/lora/lora.py,sha256=-o2mBmUvoVpdkgdAkWTARN4kfyep3UNEJLcg6moh0SU,15056
 sglang/srt/lora/lora_config.py,sha256=a2fTQESlCbG1xLiBYy4ptZ6c0Burcqyg1_6V1XSok-Y,1506
 sglang/srt/lora/lora_manager.py,sha256=DHiqdl0_4wQ5PxZBZtlCpP14515mDV2_H9tzL3Rdss8,12886
@@ -96,44 +99,44 @@ sglang/srt/managers/data_parallel_controller.py,sha256=psI4FAuBGjtdnEuwagnGdtRqv
 sglang/srt/managers/detokenizer_manager.py,sha256=nZkbwt4yty_oy8rvg4T7PbgyVLoBLohvHl25xlQpBoo,8439
 sglang/srt/managers/image_processor.py,sha256=Y8RgyrzbJjJTpjbnZDa5qiiG5wWjZ68rOXUPDi6kkFo,13698
 sglang/srt/managers/io_struct.py,sha256=_LWWqT3LNwZGaWhg2d3kTg1V2MTHKzRasCvxF9Nfpi4,15429
-sglang/srt/managers/schedule_batch.py,sha256=SAd7sxhoC3Bp8_xd-TEcXEFZBlGZPbn8-wMvBcjU55Q,45607
-sglang/srt/managers/schedule_policy.py,sha256=cLNi__smbg02keWgUMfB_nEM3vllocPB0XyG1P5qO7I,15469
-sglang/srt/managers/scheduler.py,sha256=QlcVMtrLlNcBOkVISdO556jrK8a4LE4ULskC0oCH2IQ,61776
+sglang/srt/managers/schedule_batch.py,sha256=qryPWCdOTFzxomDa80U-5guShOb1K4kBUWcPCCchYB8,45762
+sglang/srt/managers/schedule_policy.py,sha256=QxjQ8-le062AMHHxool6CxkhvB4FIwhOQPzTX_JwL6U,15447
+sglang/srt/managers/scheduler.py,sha256=Yh15uQFhJlku8a20-lhtIsiEHAcUmpL3BzL42kLVwiI,61637
 sglang/srt/managers/session_controller.py,sha256=Yp-IV3rXczACZxZXmF-QxW9CWICGy8KHQ9ttBGJ8WXA,2800
-sglang/srt/managers/tokenizer_manager.py,sha256=gnCCdB5XDobOoBKptwv-o0yYqDkMUxL78s0zBno5lM4,31219
+sglang/srt/managers/tokenizer_manager.py,sha256=uKiTt__lCFXG60zQhmM_K7dU7IuedVSIQHVw3x3y5-E,31758
 sglang/srt/managers/tp_worker.py,sha256=X1EwFX3FSsmXx7jeeX2tjZRocaujabQYWm-M-0CFEBE,7363
 sglang/srt/managers/tp_worker_overlap_thread.py,sha256=-QNBJRKxraa9Xt2WI1AFzZYdneIJ1eXv0GjFzDqXoE0,8926
 sglang/srt/mem_cache/base_prefix_cache.py,sha256=QC8HS8RC5DXu14kyXsxAgEUsn0f932p2DjqzbKjc6Bs,962
 sglang/srt/mem_cache/chunk_cache.py,sha256=R2gHAuqKd5ayQW3NnsgoGUH31---Z5izCDyCqLL0FjQ,2524
 sglang/srt/mem_cache/flush_cache.py,sha256=GYcxmNXh4hsMpFfNOuCTpKilW7guZwTtAg_usVeM3J0,979
-sglang/srt/mem_cache/memory_pool.py,sha256=l9_srwXEfIIDF46nxykbHIOo1VSvU5_Ew3H0r5EC7Fo,11072
+sglang/srt/mem_cache/memory_pool.py,sha256=oxk3UtiiFA3_1iIP6eFsk8HIcRI_8Z1-FE2KOWDr-YM,11366
 sglang/srt/mem_cache/radix_cache.py,sha256=c5voySV5L855c0G9cBEc9iQ4nR7PDDmg0V6fWWJHcq4,10945
 sglang/srt/metrics/collector.py,sha256=ZWoFx_FKN0sNMSZ8RJWUVQ0RFEYhIHxdw0d4TZTluMU,6861
 sglang/srt/metrics/func_timer.py,sha256=VFyNRrbnKVCwnQsrlLin1lITJfjQpf9m8sGPqL5LIsQ,3438
-sglang/srt/model_executor/cuda_graph_runner.py,sha256=kZ3nV03MD8EQYQB38u4_88_wyW4unECxAdMVICpPyuk,16241
+sglang/srt/model_executor/cuda_graph_runner.py,sha256=1n5WxoE9-0B3unwkkcR355K_D290h2LGt_7EvH02DQM,16246
 sglang/srt/model_executor/forward_batch_info.py,sha256=L5mVoW5SaO6To-7nGk0TZM-FFB5_78cARpJ-aC2rwD0,12883
 sglang/srt/model_executor/model_runner.py,sha256=MLYBcYIQihu2I3PBTUghiU2mSWsDMzlKzcnX7yHa9JU,29837
 sglang/srt/model_loader/__init__.py,sha256=zGZkOBz1zx-pkaIy47BasL3fjDlAcxAXUTjInOhXHAE,919
-sglang/srt/model_loader/loader.py,sha256=VBrY4W9CiVvS_D8yXhdkW9jReV9rSMSkJplabz0Fxgk,43528
+sglang/srt/model_loader/loader.py,sha256=7OG_8-66vFDFZ9kVKGNK1BFBjZ6ql449dlyvdCbMqvE,43876
 sglang/srt/model_loader/utils.py,sha256=0NaMR67fESFopaklmsleiL27XH1QUrjZW246MUu1EJ0,1369
 sglang/srt/model_loader/weight_utils.py,sha256=kQo9KPThjH3HAOCfC_tdwdrshdWuWJOVpPR0skSyaRY,24193
 sglang/srt/models/baichuan.py,sha256=PzBOFcEAixakPEkQSaJwC0Xc1fu-yCsN9T0I67r8QmY,14919
 sglang/srt/models/chatglm.py,sha256=DOrEhmb0s-yPId88R6nJeLOTUEtogk-vkB69qT2JdWc,12913
 sglang/srt/models/commandr.py,sha256=PNXgfOZF84h-rSH0edEECUmEGW8YLb44V75Z_oDhFiA,14223
-sglang/srt/models/dbrx.py,sha256=2Wqcf3sv57l4gi2xH8yrb5WSmY-4_kbbf6fhpJ4aKWw,14581
-sglang/srt/models/deepseek.py,sha256=BVNICGoLjQoHmR5lc31YrZ6YbxSRTBilHqlLsALr2u8,15693
-sglang/srt/models/deepseek_v2.py,sha256=YKSrqagVcSUwCAi-rwIph-Xu12GrNETMNKxgnffWod8,35349
+sglang/srt/models/dbrx.py,sha256=okIpIwdr8Cfrz_thzc1F75XqCUfHhFLvZ1B6BaswKoA,14585
+sglang/srt/models/deepseek.py,sha256=_cVOvR6eSEgRf6TUBpTD5uMdijDWFw4sSt4lGzl8tbg,15697
+sglang/srt/models/deepseek_v2.py,sha256=-v_OJr2c3gJ0NMxQjvT3Jknz1XPGkzKx0TVR3NIiC6A,37284
 sglang/srt/models/exaone.py,sha256=dkERTZVxrRroqu5AGLP7D4N6n8HvDqlNaDQUIe15mZY,13038
 sglang/srt/models/gemma.py,sha256=ydRqsG-7004r1fAiz01LHUmcj_6XN0Tn4xO1keJnMQk,12126
-sglang/srt/models/gemma2.py,sha256=41PlW8pMb4rMETdAni_JWDhZeIn_QsTQireAyUjsURA,15848
+sglang/srt/models/gemma2.py,sha256=-bFN-Te3YWAunLCrF-XFk_6fJS7gHM4Ca6h6aesXUTM,16362
 sglang/srt/models/gemma2_reward.py,sha256=nJ01KfqLSJtqMLm3sG8p2mGZFK1xhhjh7I7Ccb-_Hq8,2494
 sglang/srt/models/gpt2.py,sha256=2je1kE09sGcaORWnJuGYAkcwwOrT9EK-KhQaoCKjCSA,9517
 sglang/srt/models/gpt_bigcode.py,sha256=tovyOdJu2x3LkzmkdFXX_iJdkxuyChIDxwgvPBy6UPo,9528
 sglang/srt/models/granite.py,sha256=AeQY9Dxd1ZnwgCYBK0vSXXiMGM-yt9iaOVf_ruOUHXw,20409
-sglang/srt/models/grok.py,sha256=UWvVEYfEoH0jGNFSbXpO66OGW5pzmIHlNKcn9gRZEoQ,15664
+sglang/srt/models/grok.py,sha256=J9lgNbFebvXgF19nfZyHwlGPlGWY_m0LgP506YvOYrU,15668
 sglang/srt/models/internlm2.py,sha256=_xcKtd6YtEFUTozaN-yUb0xbSYckRpomfPSKcAk4j-Y,12127
 sglang/srt/models/internlm2_reward.py,sha256=8K26A9oIFFGx_9U2mF87j7FX8K87HGKMnVL3ht1Uc7I,2398
-sglang/srt/models/llama.py,sha256=S7nS05hhFGghXu0v-w9RZyBTY6OCEVF5Aaw4GX_E_9g,19929
+sglang/srt/models/llama.py,sha256=o3FYyOhkZJirzugyYz1kxs6RpY84O_uKowWWmt3jv24,19929
 sglang/srt/models/llama_classification.py,sha256=DwboM1xHXdf3Fddf7xGnrfdOLJwXdiJs994cIpAPa2g,2984
 sglang/srt/models/llama_embedding.py,sha256=rh-AiczPY_pTpzcACHvSMVjh1hsV_MZBBwP0LQxPsGM,3130
 sglang/srt/models/llama_reward.py,sha256=oPxh5E2UkxLULNdR68dFvt2I7j33CJFN6nyA-8L2_cg,4516
@@ -142,27 +145,27 @@ sglang/srt/models/llavavid.py,sha256=dYUkKfHoE15vF_VXA_s_ICCTUMSmSgvP181fk8dUi0g
 sglang/srt/models/minicpm.py,sha256=ws4AqhOfAvYHGd04QuXCZel-Oxy9_vN4p4rTjs9RSz0,13723
 sglang/srt/models/minicpm3.py,sha256=YIKJDTpwjmpLlv1sNT93k2yZMvGQlI_H87czjf6QYyo,24707
 sglang/srt/models/mistral.py,sha256=EYifJUUzN2Z2-iL37eJiNZF_DB0H4pa0mKlgYRIxM70,838
-sglang/srt/models/mixtral.py,sha256=vi6ssY75kNLy_kJrDru6gJYiAogHjSniaO6aMFd1w4E,14515
+sglang/srt/models/mixtral.py,sha256=L2Gz-Cmih1V75Ks9jmI2a6rUQ1Cl6F2uDgrhDjjDJzs,14523
 sglang/srt/models/mixtral_quant.py,sha256=uuVO1nWUZJiDhbqZN6gzSMwyfpyZorMuFXHeMCGo7N0,14022
 sglang/srt/models/mllama.py,sha256=3kX-UqeTSYZL5kPNdkfKEAEv3DpSAW1ArAAoeiXVzIc,37739
 sglang/srt/models/olmo.py,sha256=OCDMtX1OI83r80mzU4FMC3Tg8cleQ-7C8Tpoe8zgzss,11708
 sglang/srt/models/olmo2.py,sha256=aC7svioN7XT5owRxPrvhvWBNMON9QXGQBWJ1KHMyXeA,13442
-sglang/srt/models/olmoe.py,sha256=Rw-3YrHWd90MZQFnmcfUQ-3wAaI0PCFKb0DIrCDND3s,15347
+sglang/srt/models/olmoe.py,sha256=LiHVGfRaC5c_BU_vVgtV9uLuDH_SC0dw1kEc61posmI,15351
 sglang/srt/models/phi3_small.py,sha256=44_my3QmgJ2N7SOkGZzEb62DXBeCVHojfmCWgkk2uCI,14802
 sglang/srt/models/qwen.py,sha256=_FKDbwaS5C07uJyyivZpBrXJVej4Ph9ivzJdzWJPxJ4,9904
 sglang/srt/models/qwen2.py,sha256=be4xgcuqNa9kBdaL7x3PjsnUky6fh5K33c_khAWSi04,12959
-sglang/srt/models/qwen2_moe.py,sha256=rYUk_vZW3ftKIIlqPvJZ1K-6oZ_PfGspixh1zm2Y8C8,16538
+sglang/srt/models/qwen2_moe.py,sha256=6xRRJxWWh1M5UFPfvhsCpY477zv-30AeSRJXsvOkgFc,16542
 sglang/srt/models/qwen2_vl.py,sha256=3EaUlTbyWOTRXA7eViK1WqmVbCFhXLIpnos49zzf-yM,26561
 sglang/srt/models/registry.py,sha256=inKh9iwOp3LFYm3nqujg-OtABClOP-ifc1stA9cZegA,3434
 sglang/srt/models/stablelm.py,sha256=iBlIkM7CQmqI25nsujWk0LLCQD7TshzUU8qzZYYrt20,11311
 sglang/srt/models/torch_native_llama.py,sha256=YeXHorFm6QfnczLXwPb5TG9a-He0uiA9RzpR1YZKGg4,18758
 sglang/srt/models/xverse.py,sha256=Oq--KqvbYu2H4TMVGEHpSnJLEwXBpxlncR9ilsQeckc,13579
-sglang/srt/models/xverse_moe.py,sha256=AawKEQw--oAl-yzwCjoaZRG7q3rdkyDiam3FS0zjf_c,15537
+sglang/srt/models/xverse_moe.py,sha256=7E60YIST4ELYwLRgjtHiLRI5Uyc7XqQTM7jQXiWaQs4,15541
 sglang/srt/models/yivl.py,sha256=88OubtuZ38Dxb2LzfV_MTPBI4wKhh4NJqFu--efbhFM,4809
-sglang/srt/openai_api/adapter.py,sha256=dvKq4O3Rhd77ad6iCtPNykgnk9PVJE-E8wHVsBAfCQQ,53927
-sglang/srt/openai_api/protocol.py,sha256=ecRNNqkhwwKZaIoJlPhtp2VTcHxBJDbNN8lrKS7uBx8,10406
+sglang/srt/openai_api/adapter.py,sha256=X0HLuNhg-chDQjcdsQIRpZijlImEwZLHum3G0JgU4Go,54834
+sglang/srt/openai_api/protocol.py,sha256=RMzeDfh2tZITjhNwB2nX68wZwQe40N6HBuVebCzEWiU,10468
 sglang/srt/sampling/sampling_batch_info.py,sha256=s--zNjk-LErZ5lMqnZ7KiuJltaziKRbQAU5qYpKIxAc,8564
-sglang/srt/sampling/sampling_params.py,sha256=n7RbBg_bS5fYhsiWa8uJYnfoXy_i5DvtTBOkuFnHDNU,5286
+sglang/srt/sampling/sampling_params.py,sha256=BkgCJAOSmQXwJrNXg26zSjKfMy0d5mMN6oHRk_ZuESI,5499
 sglang/srt/sampling/penaltylib/__init__.py,sha256=5vQw0Y5DSzmsoFg1IdMIKLwFVhYZ5ArADHVBYbSmOec,513
 sglang/srt/sampling/penaltylib/orchestrator.py,sha256=J-DEemZcKm1--o37kf3qDOE8SZ_6H3d5oex49Mgq2ZU,10762
 sglang/srt/sampling/penaltylib/penalizers/frequency_penalty.py,sha256=1Zp2aL6dD60mwD1tCcSG0x5IYo0v4z9ce-q_YwbJ9f8,2490
@@ -180,12 +183,13 @@ sglang/test/simple_eval_math.py,sha256=6kGKNwNbLN-Af3Wj8WTimWhH-Xp3enDmSvvSjsgWU
 sglang/test/simple_eval_mgsm.py,sha256=rd7TSUyxdKbrXaVoewo24V8lCo_6kO8zxPhhmvylpw8,10259
 sglang/test/simple_eval_mmlu.py,sha256=FkwamjGMjueTixymkedF-YiPloSLiy4ftILFUrKZ9XI,4357
 sglang/test/test_activation.py,sha256=jkdNRzJnbd5OgZliQaIXpxovlcky17UrweomcOcMxoE,1442
+sglang/test/test_block_fp8.py,sha256=rhrIun8aW5zq2qvuGRlo7F7aZ_upjVxtQMVlyc2Th_E,11771
 sglang/test/test_layernorm.py,sha256=IacByD5d-stXjzBz8Ypamc7povlcedpKPbb_4JLgo3c,3720
 sglang/test/test_programs.py,sha256=1Z0umrsUu9pagzyGH5SrXl_qhKSyTfUv_kWC2mcn0qo,18208
 sglang/test/test_utils.py,sha256=HJG7kUQOk6n9FBbH89PDtQ41C3kt1cfJODhAEcFT0AQ,23823
 sglang/test/srt/sampling/penaltylib/utils.py,sha256=CjxHgywh0hx_87iynzQt_ztHu6zBVuE-YrZ-XPmW6U4,12906
-sglang-0.4.0.post2.dist-info/LICENSE,sha256=FJXh51fvTQklojUFY89XVLsjxRcBqOxPs8XNy-2uZ0c,11346
-sglang-0.4.0.post2.dist-info/METADATA,sha256=maHXecD3U1DdhzfU2aBMhN96MQRqCBPsIA1KlO7t7dg,22512
-sglang-0.4.0.post2.dist-info/WHEEL,sha256=PZUExdf71Ui_so67QXpySuHtCi3-J3wvF4ORK6k_S8U,91
-sglang-0.4.0.post2.dist-info/top_level.txt,sha256=yxhh3pYQkcnA7v3Bg889C2jZhvtJdEincysO7PEB09M,7
-sglang-0.4.0.post2.dist-info/RECORD,,
+sglang-0.4.1.post1.dist-info/LICENSE,sha256=FJXh51fvTQklojUFY89XVLsjxRcBqOxPs8XNy-2uZ0c,11346
+sglang-0.4.1.post1.dist-info/METADATA,sha256=R2YDOrUU_49x5TEbNUODNlXvkSIzFqT7-hvInlSCs5k,22527
+sglang-0.4.1.post1.dist-info/WHEEL,sha256=PZUExdf71Ui_so67QXpySuHtCi3-J3wvF4ORK6k_S8U,91
+sglang-0.4.1.post1.dist-info/top_level.txt,sha256=yxhh3pYQkcnA7v3Bg889C2jZhvtJdEincysO7PEB09M,7
+sglang-0.4.1.post1.dist-info/RECORD,,