sglang 0.4.0__tar.gz → 0.4.0.post1__tar.gz

{sglang-0.4.0 → sglang-0.4.0.post1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: sglang
-Version: 0.4.0
+Version: 0.4.0.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
@@ -239,7 +239,7 @@ Requires-Dist: xgrammar>=0.1.4; extra == "runtime-common"
 Provides-Extra: srt
 Requires-Dist: sglang[runtime_common]; extra == "srt"
 Requires-Dist: torch; extra == "srt"
-Requires-Dist: vllm>=0.6.3.post1; 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"
 Provides-Extra: srt-hip
@@ -315,6 +315,7 @@ Requires-Dist: sglang[test]; extra == "dev-hpu"
 [**Join Bi-Weekly Development Meeting**](https://docs.google.com/document/d/1xEow4eIM152xNcRxqZz9VEcOiTQo8-CEuuQ5qTmkt-E/edit?usp=sharing) | [**Slides**](https://github.com/sgl-project/sgl-learning-materials?tab=readme-ov-file#slides) |
 
 ## News
+- [2024/12] 🔥 SGLang v0.4: Zero-Overhead Batch Scheduler, Cache-Aware Load Balancer, Faster Structured Outputs ([blog](https://lmsys.org/blog/2024-12-04-sglang-v0-4/)).
 - [2024/10] 🔥 The First SGLang Online Meetup ([slides](https://github.com/sgl-project/sgl-learning-materials?tab=readme-ov-file#the-first-sglang-online-meetup)).
 - [2024/09] SGLang v0.3 Release: 7x Faster DeepSeek MLA, 1.5x Faster torch.compile, Multi-Image/Video LLaVA-OneVision ([blog](https://lmsys.org/blog/2024-09-04-sglang-v0-3/)).
 - [2024/07] Faster Llama3 Serving with SGLang Runtime (vs. TensorRT-LLM, vLLM) ([blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/)).
@@ -346,13 +347,13 @@ The core features include:
 - [Frontend: Structured Generation Language (SGLang)](https://sgl-project.github.io/frontend/frontend.html)
 
 ## Benchmark And Performance
-Learn more in our release blogs: [v0.2 blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/), [v0.3 blog](https://lmsys.org/blog/2024-09-04-sglang-v0-3/)
+Learn more in our release blogs: [v0.2 blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/), [v0.3 blog](https://lmsys.org/blog/2024-09-04-sglang-v0-3/), [v0.4 blog](https://lmsys.org/blog/2024-12-04-sglang-v0-4/)
 
 ## Roadmap
 [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, NVIDIA, RunPod, Stanford, UC Berkeley, xAI and 01.AI.
+The project is supported by (alphabetically): AMD, Baseten, Etched, Hyperbolic, Jam & Tea Studios, LinkedIn, Meituan, NVIDIA, RunPod, Stanford, UC Berkeley, xAI and 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).

{sglang-0.4.0 → sglang-0.4.0.post1}/README.md

@@ -16,6 +16,7 @@
 [**Join Bi-Weekly Development Meeting**](https://docs.google.com/document/d/1xEow4eIM152xNcRxqZz9VEcOiTQo8-CEuuQ5qTmkt-E/edit?usp=sharing) | [**Slides**](https://github.com/sgl-project/sgl-learning-materials?tab=readme-ov-file#slides) |
 
 ## News
+- [2024/12] 🔥 SGLang v0.4: Zero-Overhead Batch Scheduler, Cache-Aware Load Balancer, Faster Structured Outputs ([blog](https://lmsys.org/blog/2024-12-04-sglang-v0-4/)).
 - [2024/10] 🔥 The First SGLang Online Meetup ([slides](https://github.com/sgl-project/sgl-learning-materials?tab=readme-ov-file#the-first-sglang-online-meetup)).
 - [2024/09] SGLang v0.3 Release: 7x Faster DeepSeek MLA, 1.5x Faster torch.compile, Multi-Image/Video LLaVA-OneVision ([blog](https://lmsys.org/blog/2024-09-04-sglang-v0-3/)).
 - [2024/07] Faster Llama3 Serving with SGLang Runtime (vs. TensorRT-LLM, vLLM) ([blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/)).
@@ -47,13 +48,13 @@ The core features include:
 - [Frontend: Structured Generation Language (SGLang)](https://sgl-project.github.io/frontend/frontend.html)
 
 ## Benchmark And Performance
-Learn more in our release blogs: [v0.2 blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/), [v0.3 blog](https://lmsys.org/blog/2024-09-04-sglang-v0-3/)
+Learn more in our release blogs: [v0.2 blog](https://lmsys.org/blog/2024-07-25-sglang-llama3/), [v0.3 blog](https://lmsys.org/blog/2024-09-04-sglang-v0-3/), [v0.4 blog](https://lmsys.org/blog/2024-12-04-sglang-v0-4/)
 
 ## Roadmap
 [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, NVIDIA, RunPod, Stanford, UC Berkeley, xAI and 01.AI.
+The project is supported by (alphabetically): AMD, Baseten, Etched, Hyperbolic, Jam & Tea Studios, LinkedIn, Meituan, NVIDIA, RunPod, Stanford, UC Berkeley, xAI and 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).

{sglang-0.4.0 → sglang-0.4.0.post1}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "sglang"
-version = "0.4.0"
+version = "0.4.0.post1"
 description = "SGLang is yet another fast serving framework for large language models and vision language models."
 readme = "README.md"
 requires-python = ">=3.8"
@@ -23,7 +23,7 @@ runtime_common = ["aiohttp", "decord", "fastapi",
     "psutil", "pydantic", "python-multipart",
     "pyzmq>=25.1.2", "torchao", "uvicorn", "uvloop",
     "xgrammar>=0.1.4"]
-srt = ["sglang[runtime_common]", "torch", "vllm>=0.6.3.post1", "cuda-python", "flashinfer>=0.1.6"]
+srt = ["sglang[runtime_common]", "torch", "vllm>=0.6.3.post1,<=0.6.4.post1", "cuda-python", "flashinfer>=0.1.6"]
 
 # HIP (Heterogeneous-computing Interface for Portability) for AMD
 # => base docker rocm/vllm-dev:20241022, not from public vllm whl

{sglang-0.4.0 → sglang-0.4.0.post1}/sglang/__init__.py

@@ -66,7 +66,7 @@ from sglang.version import __version__
 
 __all__ += ["__version__"]
 
-# SGL Backends
+# SGLang Backends
 from sglang.lang.backend.runtime_endpoint import RuntimeEndpoint
 from sglang.utils import LazyImport
 

{sglang-0.4.0 → sglang-0.4.0.post1}/sglang/srt/constrained/outlines_backend.py

@@ -42,6 +42,7 @@ class OutlinesGrammar(BaseGrammarObject):
         self.guide = guide
         self.jump_forward_map = jump_forward_map
         self.state = 0
+        self.finished = False
 
     def accept_token(self, token: int):
         self.state = self.guide.get_next_state(self.state, token)
@@ -84,6 +85,10 @@ class OutlinesGrammar(BaseGrammarObject):
     ) -> torch.Tensor:
         return torch.zeros(batch_size, vocab_size, dtype=torch.bool, device=device)
 
+    @staticmethod
+    def move_vocab_mask(vocab_mask: torch.Tensor, device) -> torch.Tensor:
+        return vocab_mask
+
     def fill_vocab_mask(self, vocab_mask: torch.Tensor, idx: int) -> None:
         tokens = torch.tensor(
             self.guide.get_next_instruction(self.state).tokens, dtype=torch.int64

{sglang-0.4.0 → sglang-0.4.0.post1}/sglang/srt/constrained/xgrammar_backend.py

@@ -45,6 +45,7 @@ class XGrammarGrammar(BaseGrammarObject):
         self.matcher = matcher
         self.vocab_size = vocab_size
         self.ctx = ctx
+        self.finished = False
 
     def accept_token(self, token: int):
         assert self.matcher.accept_token(token)
@@ -85,12 +86,11 @@ class XGrammarGrammar(BaseGrammarObject):
         self.matcher.fill_next_token_bitmask(vocab_mask, idx)
 
     @staticmethod
-    def apply_vocab_mask(logits: torch.Tensor, vocab_mask: torch.Tensor) -> None:
-        if vocab_mask.device.type != logits.device.type:
-            # vocab_mask must then be on the same device as logits
-            # when applying the token bitmask, so we check and move if needed
-            vocab_mask = vocab_mask.to(logits.device)
+    def move_vocab_mask(vocab_mask: torch.Tensor, device) -> torch.Tensor:
+        return vocab_mask.to(device, non_blocking=True)
 
+    @staticmethod
+    def apply_vocab_mask(logits: torch.Tensor, vocab_mask: torch.Tensor) -> None:
         apply_token_bitmask_inplace(logits, vocab_mask)
 
     def copy(self):

{sglang-0.4.0 → sglang-0.4.0.post1}/sglang/srt/layers/attention/__init__.py

@@ -52,12 +52,13 @@ class AttentionBackend(ABC):
         v: torch.Tensor,
         layer: RadixAttention,
         forward_batch: ForwardBatch,
+        save_kv_cache: bool = True,
     ):
         """Run forward on an attention layer."""
         if forward_batch.forward_mode.is_decode():
-            return self.forward_decode(q, k, v, layer, forward_batch)
+            return self.forward_decode(q, k, v, layer, forward_batch, save_kv_cache)
         else:
-            return self.forward_extend(q, k, v, layer, forward_batch)
+            return self.forward_extend(q, k, v, layer, forward_batch, save_kv_cache)
 
     def forward_decode(
         self,
@@ -66,6 +67,7 @@ class AttentionBackend(ABC):
         v: torch.Tensor,
         layer: RadixAttention,
         forward_batch: ForwardBatch,
+        save_kv_cache: bool = True,
     ):
         """Run a forward for decode."""
         raise NotImplementedError()
@@ -77,6 +79,7 @@ class AttentionBackend(ABC):
         v: torch.Tensor,
         layer: RadixAttention,
         forward_batch: ForwardBatch,
+        save_kv_cache: bool = True,
     ):
         """Run a forward for extend."""
         raise NotImplementedError()

{sglang-0.4.0 → sglang-0.4.0.post1}/sglang/srt/layers/attention/double_sparsity_backend.py

@@ -165,7 +165,13 @@ class DoubleSparseAttnBackend(AttentionBackend):
         return 1
 
     def forward_extend(
-        self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
+        self,
+        q,
+        k,
+        v,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+        save_kv_cache=True,
     ):
         # TODO: reuse the buffer across layers
         if layer.qk_head_dim != layer.v_head_dim:
@@ -181,9 +187,10 @@ class DoubleSparseAttnBackend(AttentionBackend):
             .expand(k.shape[0], -1, -1),
         )
 
-        forward_batch.token_to_kv_pool.set_kv_buffer(
-            layer, forward_batch.out_cache_loc, k, v, k_label
-        )
+        if save_kv_cache:
+            forward_batch.token_to_kv_pool.set_kv_buffer(
+                layer, forward_batch.out_cache_loc, k, v, k_label
+            )
 
         (
             start_loc,
@@ -212,7 +219,13 @@ class DoubleSparseAttnBackend(AttentionBackend):
         return o
 
     def forward_decode(
-        self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
+        self,
+        q,
+        k,
+        v,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+        save_kv_cache=True,
     ):
         # During torch.compile, there is a bug in rotary_emb that causes the
         # output value to have a 3D tensor shape. This reshapes the output correctly.
@@ -242,9 +255,10 @@ class DoubleSparseAttnBackend(AttentionBackend):
             .expand(k.shape[0], -1, -1),
         )
 
-        forward_batch.token_to_kv_pool.set_kv_buffer(
-            layer, forward_batch.out_cache_loc, k, v, k_label
-        )
+        if save_kv_cache:
+            forward_batch.token_to_kv_pool.set_kv_buffer(
+                layer, forward_batch.out_cache_loc, k, v, k_label
+            )
 
         # NOTE(Andy) shouldn't be used when max_len_in_batch < heavy_token_num
         #            and set a minimum value for sparse_decode

{sglang-0.4.0 → sglang-0.4.0.post1}/sglang/srt/layers/attention/flashinfer_backend.py

@@ -221,7 +221,13 @@ class FlashInferAttnBackend(AttentionBackend):
         return 0
 
     def forward_extend(
-        self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
+        self,
+        q,
+        k,
+        v,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+        save_kv_cache=True,
     ):
         prefill_wrapper_paged = self.prefill_wrappers_paged[
             self._get_wrapper_idx(layer)
@@ -237,7 +243,8 @@ class FlashInferAttnBackend(AttentionBackend):
         if not use_ragged:
             if k is not None:
                 assert v is not None
-                forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
+                if save_kv_cache:
+                    forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
 
             o = prefill_wrapper_paged.forward(
                 q.contiguous().view(-1, layer.tp_q_head_num, layer.head_dim),
@@ -270,12 +277,19 @@ class FlashInferAttnBackend(AttentionBackend):
 
                 o, _ = merge_state(o1, s1, o2, s2)
 
-            forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
+            if save_kv_cache:
+                forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
 
         return o.view(-1, layer.tp_q_head_num * layer.head_dim)
 
     def forward_decode(
-        self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
+        self,
+        q,
+        k,
+        v,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+        save_kv_cache=True,
     ):
         decode_wrapper = self.forward_metadata[0][self._get_wrapper_idx(layer)]
         cache_loc = (
@@ -286,7 +300,8 @@ class FlashInferAttnBackend(AttentionBackend):
 
         if k is not None:
             assert v is not None
-            forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
+            if save_kv_cache:
+                forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
 
         o = decode_wrapper.forward(
             q.contiguous().view(-1, layer.tp_q_head_num, layer.head_dim),

{sglang-0.4.0 → sglang-0.4.0.post1}/sglang/srt/layers/attention/torch_native_backend.py

@@ -216,16 +216,23 @@ class TorchNativeAttnBackend(AttentionBackend):
         return output
 
     def forward_extend(
-        self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
+        self,
+        q,
+        k,
+        v,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+        save_kv_cache=True,
     ):
         if layer.qk_head_dim != layer.v_head_dim:
             o = q.new_empty((q.shape[0], layer.tp_q_head_num * layer.v_head_dim))
         else:
             o = torch.empty_like(q)
 
-        forward_batch.token_to_kv_pool.set_kv_buffer(
-            layer, forward_batch.out_cache_loc, k, v
-        )
+        if save_kv_cache:
+            forward_batch.token_to_kv_pool.set_kv_buffer(
+                layer, forward_batch.out_cache_loc, k, v
+            )
 
         use_gqa = layer.tp_q_head_num != layer.tp_k_head_num
 
@@ -249,7 +256,13 @@ class TorchNativeAttnBackend(AttentionBackend):
         return o
 
     def forward_decode(
-        self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
+        self,
+        q,
+        k,
+        v,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+        save_kv_cache=True,
     ):
         # During torch.compile, there is a bug in rotary_emb that causes the
         # output value to have a 3D tensor shape. This reshapes the output correctly.
@@ -260,9 +273,10 @@ class TorchNativeAttnBackend(AttentionBackend):
         else:
             o = torch.empty_like(q)
 
-        forward_batch.token_to_kv_pool.set_kv_buffer(
-            layer, forward_batch.out_cache_loc, k, v
-        )
+        if save_kv_cache:
+            forward_batch.token_to_kv_pool.set_kv_buffer(
+                layer, forward_batch.out_cache_loc, k, v
+            )
 
         use_gqa = layer.tp_q_head_num != layer.tp_k_head_num
 

{sglang-0.4.0 → sglang-0.4.0.post1}/sglang/srt/layers/attention/triton_backend.py

@@ -114,7 +114,13 @@ class TritonAttnBackend(AttentionBackend):
         return 1
 
     def forward_extend(
-        self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
+        self,
+        q,
+        k,
+        v,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+        save_kv_cache=True,
     ):
         # TODO: reuse the buffer across layers
         if layer.qk_head_dim != layer.v_head_dim:
@@ -122,9 +128,10 @@ class TritonAttnBackend(AttentionBackend):
         else:
             o = torch.empty_like(q)
 
-        forward_batch.token_to_kv_pool.set_kv_buffer(
-            layer, forward_batch.out_cache_loc, k, v
-        )
+        if save_kv_cache:
+            forward_batch.token_to_kv_pool.set_kv_buffer(
+                layer, forward_batch.out_cache_loc, k, v
+            )
 
         start_loc, attn_logits, max_seq_len, max_extend_len = self.forward_metadata
         self.extend_attention_fwd(
@@ -146,7 +153,13 @@ class TritonAttnBackend(AttentionBackend):
         return o
 
     def forward_decode(
-        self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
+        self,
+        q,
+        k,
+        v,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+        save_kv_cache=True,
     ):
         # During torch.compile, there is a bug in rotary_emb that causes the
         # output value to have a 3D tensor shape. This reshapes the output correctly.
@@ -160,9 +173,10 @@ class TritonAttnBackend(AttentionBackend):
 
         start_loc, attn_logits, max_seq_len, max_extend_len = self.forward_metadata
 
-        forward_batch.token_to_kv_pool.set_kv_buffer(
-            layer, forward_batch.out_cache_loc, k, v
-        )
+        if save_kv_cache:
+            forward_batch.token_to_kv_pool.set_kv_buffer(
+                layer, forward_batch.out_cache_loc, k, v
+            )
 
         self.decode_attention_fwd(
             q.view(-1, layer.tp_q_head_num, layer.qk_head_dim),

{sglang-0.4.0 → sglang-0.4.0.post1}/sglang/srt/layers/attention/triton_ops/extend_attention.py

@@ -284,6 +284,9 @@ def extend_attention_fwd(
     elif Lq == 288:
         BLOCK_DMODEL = 256
         BLOCK_DPE = 32
+    elif Lq == 192:
+        BLOCK_DMODEL = 128
+        BLOCK_DPE = 64
     else:
         BLOCK_DMODEL = triton.next_power_of_2(Lq)
         BLOCK_DPE = 0