sglang 0.4.1.post3__tar.gz → 0.4.1.post5__tar.gz

{sglang-0.4.1.post3 → sglang-0.4.1.post5}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.1
+Metadata-Version: 2.2
 Name: sglang
-Version: 0.4.1.post3
+Version: 0.4.1.post5
 Summary: SGLang is yet another fast serving framework for large language models and vision language models.
 License:                                  Apache License
                                    Version 2.0, January 2004
@@ -239,15 +239,15 @@ Requires-Dist: uvloop; extra == "runtime-common"
 Requires-Dist: xgrammar>=0.1.6; extra == "runtime-common"
 Provides-Extra: srt
 Requires-Dist: sglang[runtime_common]; extra == "srt"
+Requires-Dist: cuda-python; extra == "srt"
+Requires-Dist: sgl-kernel>=0.0.2.post11; extra == "srt"
 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"
-Requires-Dist: vllm==0.6.3.dev13; extra == "srt-hip"
+Requires-Dist: vllm==0.6.3.post2.dev1; extra == "srt-hip"
 Provides-Extra: srt-xpu
 Requires-Dist: sglang[runtime_common]; extra == "srt-xpu"
 Provides-Extra: srt-hpu
@@ -315,7 +315,7 @@ Requires-Dist: sglang[test]; extra == "dev-hpu"
 
 | [**Blog**](https://lmsys.org/blog/2024-07-25-sglang-llama3/)
 | [**Documentation**](https://sgl-project.github.io/)
-| [**Join Slack**](https://join.slack.com/t/sgl-fru7574/shared_invite/zt-2tmmp6flg-89dOlJW2TjnBrTRk1I_~GA)
+| [**Join Slack**](https://join.slack.com/t/sgl-fru7574/shared_invite/zt-2um0ad92q-LkU19KQTxCGzlCgRiOiQEw)
 | [**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) |
 
@@ -347,12 +347,13 @@ The core features include:
 
 ## Getting Started
 - [Install SGLang](https://sgl-project.github.io/start/install.html)
-- [Send requests](https://sgl-project.github.io/start/send_request.html)
-- [Backend: SGLang Runtime (SRT)](https://sgl-project.github.io/backend/backend.html)
-- [Frontend: Structured Generation Language (SGLang)](https://sgl-project.github.io/frontend/frontend.html)
+- [Quick Start](https://sgl-project.github.io/start/send_request.html)
+- [Backend Tutorial](https://sgl-project.github.io/backend/openai_api_completions.html)
+- [Frontend Tutorial](https://sgl-project.github.io/frontend/frontend.html)
+- [Contribution Guide](https://sgl-project.github.io/references/contribution_guide.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/), [v0.4 blog](https://lmsys.org/blog/2024-12-04-sglang-v0-4/)
+Learn more in the 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)
@@ -361,5 +362,4 @@ Learn more in our release blogs: [v0.2 blog](https://lmsys.org/blog/2024-07-25-s
 The project is supported by (alphabetically): AMD, Baseten, DataCrunch, Etched, Hyperbolic, Jam & Tea Studios, LinkedIn, LMSYS.org, Meituan, NVIDIA, RunPod, Stanford, UC Berkeley, UCLA, 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 the paper, [SGLang: Efficient Execution of Structured Language Model Programs](https://arxiv.org/abs/2312.07104), if you find the project useful.
+We learned 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 the paper, [SGLang: Efficient Execution of Structured Language Model Programs](https://arxiv.org/abs/2312.07104), if you find the project useful.

{sglang-0.4.1.post3 → sglang-0.4.1.post5}/README.md

@@ -14,7 +14,7 @@
 
 | [**Blog**](https://lmsys.org/blog/2024-07-25-sglang-llama3/)
 | [**Documentation**](https://sgl-project.github.io/)
-| [**Join Slack**](https://join.slack.com/t/sgl-fru7574/shared_invite/zt-2tmmp6flg-89dOlJW2TjnBrTRk1I_~GA)
+| [**Join Slack**](https://join.slack.com/t/sgl-fru7574/shared_invite/zt-2um0ad92q-LkU19KQTxCGzlCgRiOiQEw)
 | [**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) |
 
@@ -46,12 +46,13 @@ The core features include:
 
 ## Getting Started
 - [Install SGLang](https://sgl-project.github.io/start/install.html)
-- [Send requests](https://sgl-project.github.io/start/send_request.html)
-- [Backend: SGLang Runtime (SRT)](https://sgl-project.github.io/backend/backend.html)
-- [Frontend: Structured Generation Language (SGLang)](https://sgl-project.github.io/frontend/frontend.html)
+- [Quick Start](https://sgl-project.github.io/start/send_request.html)
+- [Backend Tutorial](https://sgl-project.github.io/backend/openai_api_completions.html)
+- [Frontend Tutorial](https://sgl-project.github.io/frontend/frontend.html)
+- [Contribution Guide](https://sgl-project.github.io/references/contribution_guide.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/), [v0.4 blog](https://lmsys.org/blog/2024-12-04-sglang-v0-4/)
+Learn more in the 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)
@@ -60,5 +61,4 @@ Learn more in our release blogs: [v0.2 blog](https://lmsys.org/blog/2024-07-25-s
 The project is supported by (alphabetically): AMD, Baseten, DataCrunch, Etched, Hyperbolic, Jam & Tea Studios, LinkedIn, LMSYS.org, Meituan, NVIDIA, RunPod, Stanford, UC Berkeley, UCLA, 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 the paper, [SGLang: Efficient Execution of Structured Language Model Programs](https://arxiv.org/abs/2312.07104), if you find the project useful.
+We learned 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 the paper, [SGLang: Efficient Execution of Structured Language Model Programs](https://arxiv.org/abs/2312.07104), if you find the project useful.

{sglang-0.4.1.post3 → sglang-0.4.1.post5}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "sglang"
-version = "0.4.1.post3"
+version = "0.4.1.post5"
 description = "SGLang is yet another fast serving framework for large language models and vision language models."
 readme = "README.md"
 requires-python = ">=3.8"
@@ -16,18 +16,24 @@ classifiers = [
 dependencies = ["requests", "tqdm", "numpy", "IPython", "setproctitle"]
 
 [project.optional-dependencies]
-runtime_common = ["aiohttp", "decord", "fastapi",
+runtime_common = [
+    "aiohttp", "decord", "fastapi",
     "hf_transfer", "huggingface_hub", "interegular", "modelscope",
     "orjson", "outlines>=0.0.44,<0.1.0",
     "packaging", "pillow", "prometheus-client>=0.20.0",
     "psutil", "pydantic", "python-multipart",
     "pyzmq>=25.1.2", "torchao>=0.7.0", "uvicorn", "uvloop",
-    "xgrammar>=0.1.6"]
-srt = ["sglang[runtime_common]", "torch", "vllm>=0.6.3.post1,<=0.6.4.post1", "cuda-python", "flashinfer==0.1.6", "sgl-kernel>=0.0.2.post10"]
+    "xgrammar>=0.1.6"
+]
+srt = [
+    "sglang[runtime_common]", "cuda-python",
+    "sgl-kernel>=0.0.2.post11", "torch", "vllm>=0.6.3.post1,<=0.6.4.post1",
+    "flashinfer==0.1.6"
+]
 
 # HIP (Heterogeneous-computing Interface for Portability) for AMD
 # => base docker rocm/vllm-dev:20241022, not from public vllm whl
-srt_hip = ["sglang[runtime_common]", "torch", "vllm==0.6.3.dev13"]
+srt_hip = ["sglang[runtime_common]", "torch", "vllm==0.6.3.post2.dev1"]
 # xpu is not enabled in public vllm and torch whl,
 # need to follow https://docs.vllm.ai/en/latest/getting_started/xpu-installation.htmlinstall vllm
 srt_xpu = ["sglang[runtime_common]"]

{sglang-0.4.1.post3 → sglang-0.4.1.post5}/sglang/bench_one_batch.py

@@ -63,6 +63,7 @@ from sglang.srt.model_executor.model_runner import ModelRunner
 from sglang.srt.sampling.sampling_params import SamplingParams
 from sglang.srt.server import _set_envs_and_config
 from sglang.srt.server_args import PortArgs, ServerArgs
+from sglang.srt.speculative.spec_info import SpeculativeAlgorithm
 from sglang.srt.utils import configure_logger, kill_process_tree, suppress_other_loggers
 
 
@@ -214,6 +215,7 @@ def extend(reqs, model_runner):
         tree_cache=None,
         model_config=model_runner.model_config,
         enable_overlap=False,
+        spec_algorithm=SpeculativeAlgorithm.NONE,
     )
     batch.prepare_for_extend()
     model_worker_batch = batch.get_model_worker_batch()

{sglang-0.4.1.post3 → sglang-0.4.1.post5}/sglang/bench_serving.py

@@ -514,6 +514,8 @@ class BenchmarkMetrics:
     p99_itl_ms: float
     mean_e2e_latency_ms: float
     median_e2e_latency_ms: float
+    std_e2e_latency_ms: float
+    p99_e2e_latency_ms: float
 
 
 SHAREGPT_URL = "https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json"
@@ -563,7 +565,7 @@ def sample_sharegpt_requests(
         raise ValueError("output_len too small")
 
     # Download sharegpt if necessary
-    if not os.path.isfile(dataset_path):
+    if not os.path.isfile(dataset_path) and dataset_path == "":
         dataset_path = download_and_cache_file(SHAREGPT_URL)
 
     # Load the dataset.
@@ -873,6 +875,8 @@ def calculate_metrics(
         p99_itl_ms=np.percentile(itls or 0, 99) * 1000,
         mean_e2e_latency_ms=np.mean(e2e_latencies) * 1000,
         median_e2e_latency_ms=np.median(e2e_latencies) * 1000,
+        std_e2e_latency_ms=np.std(e2e_latencies) * 1000,
+        p99_e2e_latency_ms=np.percentile(e2e_latencies, 99) * 1000,
     )
 
     return metrics, output_lens
@@ -1064,8 +1068,21 @@ async def benchmark(
             "total_output_tokens_retokenized": metrics.total_output_retokenized,
             "mean_e2e_latency_ms": metrics.mean_e2e_latency_ms,
             "median_e2e_latency_ms": metrics.median_e2e_latency_ms,
+            "std_e2e_latency_ms": metrics.std_e2e_latency_ms,
+            "p99_e2e_latency_ms": metrics.p99_e2e_latency_ms,
+            "mean_ttft_ms": metrics.mean_ttft_ms,
             "median_ttft_ms": metrics.median_ttft_ms,
+            "std_ttft_ms": metrics.std_ttft_ms,
+            "p99_ttft_ms": metrics.p99_ttft_ms,
+            "mean_tpot_ms": metrics.mean_tpot_ms,
+            "median_tpot_ms": metrics.median_tpot_ms,
+            "std_tpot_ms": metrics.std_tpot_ms,
+            "p99_tpot_ms": metrics.p99_tpot_ms,
+            "mean_itl_ms": metrics.mean_itl_ms,
             "median_itl_ms": metrics.median_itl_ms,
+            "std_itl_ms": metrics.std_itl_ms,
+            "p99_itl_ms": metrics.p99_itl_ms,
+            "input_throughput": metrics.input_throughput,
             "output_throughput": metrics.output_throughput,
             "sharegpt_output_len": args.sharegpt_output_len,
             "random_input_len": args.random_input_len,

{sglang-0.4.1.post3 → sglang-0.4.1.post5}/sglang/lang/interpreter.py

@@ -96,6 +96,7 @@ def run_program_batch(
     default_sampling_para,
     num_threads,
     progress_bar,
+    generator_style=False,
 ):
     if hasattr(backend, "endpoint"):
         backend = backend.endpoint
@@ -109,6 +110,17 @@ def run_program_batch(
         num_threads = max(96, multiprocessing.cpu_count() * 16)
     num_threads = min(num_threads, len(batch_arguments))
 
+    if generator_style:
+        return _run_program_batch_generator(
+            program,
+            backend,
+            batch_arguments,
+            default_sampling_para,
+            num_threads,
+            progress_bar,
+        )
+
+    # Original code path when generator_style=False
     if num_threads == 1:
         rets = []
         if progress_bar:
@@ -168,6 +180,64 @@ def run_program_batch(
     return rets
 
 
+def _run_program_batch_generator(
+    program,
+    backend,
+    batch_arguments,
+    default_sampling_para,
+    num_threads,
+    progress_bar,
+):
+    """Helper function that yields results one by one using chunking to avoid overwhelming ThreadPoolExecutor."""
+    if num_threads == 1:
+        iterator = tqdm.tqdm(batch_arguments) if progress_bar else batch_arguments
+        for arguments in iterator:
+            yield run_program(
+                program,
+                backend,
+                (),
+                arguments,
+                default_sampling_para,
+                False,
+                True,
+            )
+    else:
+        pbar = tqdm.tqdm(total=len(batch_arguments)) if progress_bar else None
+
+        # Process in chunks to avoid overwhelming ThreadPoolExecutor
+        # Otherwise, ThreadPoolExecutor.submit will block after adding certain number of tasks
+        # so we will never reach "yield" until all tasks are done
+        chunk_size = 200
+
+        with ThreadPoolExecutor(num_threads) as executor:
+            for chunk_start in range(0, len(batch_arguments), chunk_size):
+                chunk_end = min(chunk_start + chunk_size, len(batch_arguments))
+                chunk_futures = []
+
+                # Submit chunk of tasks
+                for i in range(chunk_start, chunk_end):
+                    future = executor.submit(
+                        run_program,
+                        program,
+                        backend,
+                        (),
+                        batch_arguments[i],
+                        default_sampling_para,
+                        False,
+                        True,
+                    )
+                    if pbar:
+                        future.add_done_callback(lambda _: pbar.update())
+                    chunk_futures.append(future)
+
+                # Yield results from this chunk as they complete
+                for future in chunk_futures:
+                    yield future.result()
+
+        if pbar:
+            pbar.close()
+
+
 def cache_program(program, backend):
     from sglang.lang.tracer import extract_prefix_by_tracing
 
@@ -277,7 +347,7 @@ class StreamExecutor:
         size: int = 1,
         position_ids_offset: Optional[List[int]] = None,
     ):
-        if size > 1:
+        if size > 1 and str(self.text_):
             self.submit(SglCommitLazy())
 
         self.sync()

{sglang-0.4.1.post3 → sglang-0.4.1.post5}/sglang/lang/ir.py

@@ -227,6 +227,7 @@ class SglFunction:
         backend=None,
         num_threads: Union[str, int] = "auto",
         progress_bar: bool = False,
+        generator_style: bool = False,
     ):
         from sglang.lang.interpreter import run_program_batch
 
@@ -277,6 +278,7 @@ class SglFunction:
             default_sampling_para,
             num_threads,
             progress_bar,
+            generator_style=generator_style,
         )
 
     def trace(self, *, backend=None, **kwargs):

{sglang-0.4.1.post3 → sglang-0.4.1.post5}/sglang/srt/configs/__init__.py

@@ -1,3 +1,5 @@
+from sglang.srt.configs.chatglm import ChatGLMConfig
+from sglang.srt.configs.dbrx import DbrxConfig
 from sglang.srt.configs.exaone import ExaoneConfig
 from sglang.srt.configs.qwen2vl import Qwen2VLConfig, Qwen2VLVisionConfig
 
@@ -5,4 +7,6 @@ __all__ = [
     "ExaoneConfig",
     "Qwen2VLConfig",
     "Qwen2VLVisionConfig",
+    "ChatGLMConfig",
+    "DbrxConfig",
 ]

sglang-0.4.1.post5/sglang/srt/configs/chatglm.py

@@ -0,0 +1,78 @@
+# Adapted from
+# https://github.com/THUDM/ChatGLM2-6B
+# https://github.com/vllm-project/vllm/blob/main/vllm/transformers_utils/configs/chatglm.py
+
+# ChatGLM2 and ChatGLM3 share the same config.
+# ChatGLM4 is officially supported by Huggingface
+# transformers >= 4.46.0 is required
+# https://huggingface.co/docs/transformers/en/model_doc/glm
+from transformers import PretrainedConfig
+
+
+class ChatGLMConfig(PretrainedConfig):
+    model_type = "chatglm"
+    attribute_map = {
+        "num_hidden_layers": "num_layers",
+        "n_head_kv": "multi_query_group_num",
+    }
+
+    def __init__(
+        self,
+        num_layers=28,
+        padded_vocab_size=65024,
+        hidden_size=4096,
+        ffn_hidden_size=13696,
+        kv_channels=128,
+        num_attention_heads=32,
+        seq_length=2048,
+        hidden_dropout=0.0,
+        attention_dropout=0.0,
+        layernorm_epsilon=1e-5,
+        rmsnorm=True,
+        apply_residual_connection_post_layernorm=False,
+        post_layer_norm=True,
+        add_bias_linear=False,
+        add_qkv_bias=False,
+        interleaved_qkv=False,
+        bias_dropout_fusion=True,
+        multi_query_attention=False,
+        multi_query_group_num=1,
+        apply_query_key_layer_scaling=True,
+        attention_softmax_in_fp32=True,
+        fp32_residual_connection=False,
+        quantization_bit=0,
+        pre_seq_len=None,
+        prefix_projection=False,
+        **kwargs
+    ):
+        self.num_layers = num_layers
+        self.vocab_size = padded_vocab_size
+        self.padded_vocab_size = padded_vocab_size
+        self.hidden_size = hidden_size
+        self.ffn_hidden_size = ffn_hidden_size
+        self.kv_channels = kv_channels
+        self.num_attention_heads = num_attention_heads
+        self.seq_length = seq_length
+        # It is to be compatible with long lora.
+        self.max_position_embeddings = seq_length
+        self.hidden_dropout = hidden_dropout
+        self.attention_dropout = attention_dropout
+        self.layernorm_epsilon = layernorm_epsilon
+        self.rmsnorm = rmsnorm
+        self.apply_residual_connection_post_layernorm = (
+            apply_residual_connection_post_layernorm
+        )
+        self.post_layer_norm = post_layer_norm
+        self.add_bias_linear = add_bias_linear
+        self.add_qkv_bias = add_qkv_bias
+        self.bias_dropout_fusion = bias_dropout_fusion
+        self.multi_query_attention = multi_query_attention
+        self.multi_query_group_num = multi_query_group_num
+        self.apply_query_key_layer_scaling = apply_query_key_layer_scaling
+        self.attention_softmax_in_fp32 = attention_softmax_in_fp32
+        self.fp32_residual_connection = fp32_residual_connection
+        self.quantization_bit = quantization_bit
+        self.pre_seq_len = pre_seq_len
+        self.prefix_projection = prefix_projection
+        self.interleaved_qkv = interleaved_qkv
+        super().__init__(**kwargs)