117 Pull requests merged by 45 people

• Split test_mla.py into two files (deepseek v2 and deepseek v3)

  #4216 merged Mar 8, 2025

• refine quant kernel code style

  #4211 merged Mar 8, 2025

• Test no vllm custom allreduce

  #4210 merged Mar 8, 2025

• Refactor Dockerfile: unify CUDA logic and reduce image size by ~2.6 GB

  #3749 merged Mar 8, 2025

• Use clang format 18 in pr-test-sgl-kernel.yml

  #4203 merged Mar 8, 2025

• Fix bench_serving flush cache not recognizing OPENAI_API_KEY

  #4181 merged Mar 8, 2025

• lazy import attn backends

  #4200 merged Mar 8, 2025

• Revert "Minor improvement to per_tensor_quant_fp8 (#4197)"

  #4198 merged Mar 8, 2025

• Minor improvement to per_tensor_quant_fp8

  #4197 merged Mar 8, 2025

• Remove the vllm dependency from the moe_align function

  #4164 merged Mar 8, 2025

• [EAGLE] many fixes for eagle

  #4195 merged Mar 8, 2025

• New clang format for sgl kernel

  #4194 merged Mar 8, 2025

• Update amd ci docker image to v0.4.3.post4-rocm630.

  #4189 merged Mar 7, 2025

• Fix eagle hang issue for max_new_tokens=1

  #4185 merged Mar 7, 2025

• use same version for ci and pyproject

  #4187 merged Mar 7, 2025

• Revert "ROCm: Flex Attention Enablement with custom backends (#4178)"

  #4186 merged Mar 7, 2025

• ROCm: Flex Attention Enablement with custom backends

  #4178 merged Mar 7, 2025

• [Docs] Improve bullets appearance and grammar

  #4174 merged Mar 7, 2025

• fix int8 doc link

  #4179 merged Mar 7, 2025

• Add an example of using deepseekv3 int8 sglang.

  #4177 merged Mar 7, 2025

• chore: bump v0.0.3.post7 for sgl-kernel

  #4176 merged Mar 7, 2025

• Memory pool fix for upstream change about eagle

  #4170 merged Mar 7, 2025

• Put utils in ifndef USE_ROCM to fix CI (#4167)

  #4168 merged Mar 7, 2025

• Put utils in ifndef USE_ROCM to fix CI

  #4167 merged Mar 7, 2025

• Remove non-existent AMD header include

  #4166 merged Mar 7, 2025

• [Docs] Fix links and grammar issues

  #4162 merged Mar 7, 2025

• [Refactor] Reducing code duplication across FP8 CUDA quantization kernels

  #4163 merged Mar 7, 2025

• [Feature] DeepSeek V3/R1 INT8 Quantization (channel-wise)

  #3888 merged Mar 7, 2025

• Add sgl_per_token_quant_fp8

  #4089 merged Mar 7, 2025

• [quant kernel] sgl-kernel support per_tensor_quant fp8

  #3786 merged Mar 7, 2025

• Add Support for Qwen2-VL Multi-modal Embedding Models

  #3694 merged Mar 7, 2025

• ROCm: enable trillion-parameter MoE models with INT4-FP8 single node

  #4152 merged Mar 6, 2025

• Hot fix small vocal eagle in docs

  #4154 merged Mar 6, 2025

• Docs: add torch compile cache

  #4151 merged Mar 6, 2025

• [docs] fix HF reference script command

  #4148 merged Mar 6, 2025

• Release v0.4.3.post4

  #4140 merged Mar 6, 2025

• Fix constrained generation errors by adding datasets dependency

  #4142 merged Mar 6, 2025

• Fix nightly ci Gsm8k & Fix flashinfer backend kvcache quant

  #4147 merged Mar 6, 2025

• Fix a draft model accuracy bug in eagle; support step=1; return logprob in eagle

  #4134 merged Mar 6, 2025

• fix bench serving bug

  #4135 merged Mar 6, 2025

• Update CODEOWNER

  #4138 merged Mar 6, 2025

• AMD/ROCm: update base image string

  #4137 merged Mar 6, 2025

• [Minor] make the __init__ function of model_runner.py shorter

  #4132 merged Mar 6, 2025

• Split the __init__ of scheduler as smaller functions. Improve the eagle tests

  #4128 merged Mar 6, 2025

• remove unused max_jobs in setup_rocm.py

  #4126 merged Mar 6, 2025

• Add tag suffix to nightly docker builds.

  #4129 merged Mar 6, 2025

• Add codeowners for eagle implementations

  #4131 merged Mar 6, 2025

• EAGLE docs

  #4038 merged Mar 6, 2025

• feat: support docs auto live-reload with sphinx-autobuild

  #4111 merged Mar 6, 2025

• Add a pointer to the real KV cache pool

  #4113 merged Mar 6, 2025

• Remove prefill-only-one-req

  #4117 merged Mar 6, 2025

• [Hoxfix] Fix incomplete token_to_kv_pool refactor

  #4121 merged Mar 6, 2025

• chore: bump v0.4.3.post3

  #4114 merged Mar 6, 2025

• fix Non-consecutive header level increase in docs/router/router.md

  #4099 merged Mar 6, 2025

• fix cross-reference error and spelling mistakes

  #4101 merged Mar 6, 2025

• Online serving benchmarks of real datasets for hierarchical KV caching

  #3211 merged Mar 6, 2025

• Debug radixcache: refactor recursive helper methods

  #3029 merged Mar 6, 2025

• remove testing on PR workflow change

  #4110 merged Mar 6, 2025

• Create release-docker-amd-nightly.yml

  #4105 merged Mar 5, 2025

• revert deepseek docs

  #4109 merged Mar 5, 2025

• Add examples for server token-in-token-out

  #4103 merged Mar 5, 2025

• reorganize dpsk docs

  #4108 merged Mar 5, 2025

• Add DeepSeek optimization ablations documentation

  #4107 merged Mar 5, 2025

• Add update_weights_from_disk endpoint to Engine

  #4102 merged Mar 5, 2025

• Fix triton kernel illegal memory issue for eagle

  #4100 merged Mar 5, 2025

• [Revision] Add fast decode plan for flashinfer mla

  #4012 merged Mar 5, 2025

• Fix the moe padding conditional logic

  #4081 merged Mar 5, 2025

• [Eagle] Refactor eagle speculative decoding

  #3986 merged Mar 5, 2025

• ROCM: AITER BLOCK GEMM

  #4075 merged Mar 5, 2025

• bench: add dataset param for bench_multiturn

  #3990 merged Mar 5, 2025

• [QUANT] Add GPTQModel Dynamic Quantization + lm_head Quantization

  #3790 merged Mar 5, 2025

• test: add vlm to token in & out example

  #3941 merged Mar 5, 2025

• [Minor] more code cleanup

  #4077 merged Mar 5, 2025

• Add examples for returning hidden states when using the server

  #4074 merged Mar 5, 2025

• Simplify eagle tests and TP sync in grammar backend

  #4066 merged Mar 4, 2025

• Update nextn ci test

  #4071 merged Mar 4, 2025

• Revert "Fix nightly-test CI"

  #4065 merged Mar 4, 2025

• [Feature] Add test for speculative_token_map

  #4016 merged Mar 4, 2025

• remove unused max_jobs

  #3607 merged Mar 4, 2025

• fix: support gelu_new activation function in gpt2

  #3712 merged Mar 4, 2025

• sgl-router - issues on routing and project build. (#3870)

  #3948 merged Mar 4, 2025

• [XCCL] Use xccl for xpu backend since xccl is ready in latest PyTorch.

  #3954 merged Mar 4, 2025

• [Fix & Style] Refactor the grammar backend to reduce human errors and improve readability

  #4030 merged Mar 4, 2025

• Remove grafana dashboard's datasource uid

  #4051 merged Mar 4, 2025

• Fix debug_tensor_dump_output_folder optional key missing

  #4046 merged Mar 4, 2025

• ROCm: update aiter and its usage to fused moe (bloat16, fp8, fp8 block-quant)

  #4053 merged Mar 4, 2025

• Fix breakage problem when using custom_ar

  #4052 merged Mar 4, 2025

• HotFix for #3988 using blockwise_int8

  #4023 merged Mar 4, 2025

• Reasoning parser

  #4000 merged Mar 4, 2025

• Fix assert options.num_stages != 0 error in the latest ROCm build image

  #4049 merged Mar 4, 2025

• docs: update README

  #4044 merged Mar 4, 2025

• Add a link to the roadmap in README.md

  #4043 merged Mar 4, 2025

• Share target model embed and head weights for nextn

  #4033 merged Mar 3, 2025

• Add examples in sampling parameters

  #4039 merged Mar 3, 2025

• Remove outdated test utils and fix links for the doc of sampling params

  #3999 merged Mar 3, 2025

• Docs: Fix sampling parameter

  #4034 merged Mar 3, 2025

• Misc clean up; Remove the support of jump forward

  #4032 merged Mar 3, 2025

• Reorganize python source files in sgl-kernel with multiple files

  #4027 merged Mar 3, 2025

• Reorganize c++ source files in sgl-kernel with multiple folders

  #4025 merged Mar 3, 2025

• Update metrics documentation

  #3264 merged Mar 3, 2025

• remove cache configs in model definitions

  #4031 merged Mar 3, 2025

• Clean up custom allreduce

  #4029 merged Mar 3, 2025

• Improve code styles

  #4021 merged Mar 3, 2025

• Support penalty in overlap mode; return logprob with chunked prefill; improve benchmark scripts

  #3988 merged Mar 3, 2025

• Optimize Triton Kernel of Group GEMM in DeepGEMM Benchmark

  #4014 merged Mar 3, 2025

• Fix nightly-test CI

  #3826 merged Mar 3, 2025

• Add examples to token-in-token-out for LLM

  #4010 merged Mar 3, 2025

• Revert "Add fast decode plan for flashinfer mla"

  #4008 merged Mar 3, 2025

• Add fast decode plan for flashinfer mla

  #3987 merged Mar 3, 2025

• [feat] add small vocab table for eagle's draft model[1].

  #3822 merged Mar 3, 2025

• Add Benchmark for DeepGEMM Group GEMM

  #3993 merged Mar 3, 2025

• Enable custom AR for AMD GPUs and maintain it in sgl-kernel

  #3406 merged Mar 2, 2025

• Add accuracy test for TP torch compile

  #3994 merged Mar 2, 2025

• Fix all gather torch compile

  #3992 merged Mar 2, 2025

• fix typo

  #3991 merged Mar 2, 2025

• add deepgemm and sglang fp8 block-wise gemm benchmark

  #3893 merged Mar 2, 2025

• Update CODEOWNERS

  #3989 merged Mar 2, 2025

35 Pull requests opened by 29 people

• Reasoning parser

  #4001 opened Mar 2, 2025

• Hierarchical Caching supports MLA

  #4009 opened Mar 3, 2025

• ROCM support tree_speculative_sampling_target_only

  #4015 opened Mar 3, 2025

• Avoid duplicated request ids in batch APIs

  #4026 opened Mar 3, 2025

• Added fix for TP>1 grammar sync for slow networks

  #4040 opened Mar 3, 2025

• chore: bump v0.4.4

  #4041 opened Mar 3, 2025

• Let `bench_one_batch` support `enable_dp_attention`

  #4058 opened Mar 4, 2025

• codegen for blackwell

  #4063 opened Mar 4, 2025

• Fix RotaryEmbedding when using Triton backend for EXAONE-3.5-2.4B

  #4064 opened Mar 4, 2025

• Tool call with text

  #4067 opened Mar 4, 2025

• [WIP] Support overlapping two batches

  #4068 opened Mar 4, 2025

• add INT8 example into dsv3 README

  #4079 opened Mar 5, 2025

• Hierarchical Caching Refactoring and Fixing TP issue

  #4082 opened Mar 5, 2025

• Enable the native path of DeepSeek

  #4086 opened Mar 5, 2025

• Add awq dequantize kernel to sgl with 1x to 3x speedup

  #4104 opened Mar 5, 2025

• Support cuda graph for LoRA

  #4115 opened Mar 6, 2025

• docs(reasoning content): :memo: deepseek-r1 parser support qwq

  #4124 opened Mar 6, 2025

• Add test for Radix cache variants

  #4125 opened Mar 6, 2025

• Add A800 tuning configs support DeepSeek V3/R1 BF16 and INT8(block-wise)

  #4136 opened Mar 6, 2025

• [QUANT] Support DeepSeek-V3 gptq

  #4139 opened Mar 6, 2025

• add --served-model-name arg for bench_serving

  #4141 opened Mar 6, 2025

• fix the input_ids is None error

  #4144 opened Mar 6, 2025

• [ROCm] Enable silu_and_mul, gelu_and_mul, gelu_tanh_and_mul in amd platform

  #4150 opened Mar 6, 2025

• DeepGemm integrate to sgl-kernel

  #4165 opened Mar 7, 2025

• [ROCm/Draft/No-Merge]: Flex Attention Enablement

  #4172 opened Mar 7, 2025

• Fix MoE quant args

  #4190 opened Mar 8, 2025

• [docs] Unhide production metrics page

  #4193 opened Mar 8, 2025

• linear support deepgemm

  #4199 opened Mar 8, 2025

• Statistical Analysis of the Output Stability of the Deepseek Model

  #4202 opened Mar 8, 2025

• Added example for multimodal embedding

  #4206 opened Mar 8, 2025

• Try sgl kernel 0.0.3.post7

  #4207 opened Mar 8, 2025

• [Fix] Check the device backend before calling empty_cache function

  #4212 opened Mar 8, 2025

• Rename files in sgl kernel to avoid nested folder structure

  #4213 opened Mar 8, 2025

• Remove vllm ops scaled fp8 quant

  #4215 opened Mar 8, 2025

• Check eagle server args

  #4217 opened Mar 8, 2025

47 Issues closed by 20 people

• [Bug] launch dsv2 service failed

  #2763 closed Mar 9, 2025

• [Bug] Flashinferv0.2.2.post1 shows Unsupported max_mma_kv: 0 error on L40 , when deploying Deepseek-V2-Lite-chat with --enable-flashinfer-mla

  #4196 closed Mar 8, 2025

• Deploy QwQ-32B with 0.4.3.post4 cannot find curand_kernel.h

  #4169 closed Mar 8, 2025

• [Bug] cannot find this project in hugging face :lmsys/sglang-ci-dsv3-test.

  #4192 closed Mar 8, 2025

• [Feature] Implement Dynamic Depth Decoding (DDD) for EAGLE-2

  #2754 closed Mar 8, 2025

• [Feature] several features for veRL integration

  #2736 closed Mar 8, 2025

• [Bug] SGLang stops working after a few requests when serving DeepSeek V3

  #2741 closed Mar 8, 2025

• Decode out of memory happened when run deepseek-r1 inference

  #4184 closed Mar 7, 2025

• [Bug] The radix cache affects the accuracy of the output results.

  #4175 closed Mar 7, 2025

• I wonder if the offline engine API supports OpenAI input format.

  #2734 closed Mar 7, 2025

• permission issue in newly updated docker lmsysorg/sglang:v0.4.3.post2-rocm630

  #4122 closed Mar 6, 2025

• [Bug] Inaccurate or Inconsistent Output in Qwen2.5-VL Multi-Image Testing with sglang

  #4123 closed Mar 6, 2025

• [Bug] self.token_to_kv_pool.available_size(): AttributeError: 'Scheduler' object has no attribute 'token_to_kv_pool' with Qwen/QwQ-32B on SGLang v0.4.3.post3

  #4127 closed Mar 6, 2025

• [Bug] sgl.Engine(**dataclasses.asdict(server_args)) return_logprob=True error

  #4085 closed Mar 6, 2025

• [Bug] pydantic validation errors for ChatCompletion

  #3637 closed Mar 6, 2025

• [Feature] Reasoning model API support

  #3043 closed Mar 6, 2025

• Request help: VRAM usage issues

  #4118 closed Mar 6, 2025

• [Bug] sglang crashes with multi node

  #3932 closed Mar 6, 2025

• How can I close the server?

  #4080 closed Mar 6, 2025

• [Kernel] Launch two kernels for mixed chunked prefill

  #2273 closed Mar 6, 2025

• [Feature] Add examples for server token-in-token-out

  #4078 closed Mar 5, 2025

• [Bug] HiCacheController gets stuck when testing with multiple long text documents

  #3998 closed Mar 5, 2025

• [Bug] Qwen2.5-32B-Instruct-GPTQ-Int answer end abnormally，Qwen2.5-32B-Instruct answer is ok，use vllm is ok

  #4013 closed Mar 5, 2025

• offline throughput benchmark numbers mismatch with custom benchmark by 3x

  #4076 closed Mar 5, 2025

• [Bug] NCCL Crash with SIGSEGV Frequently when deploying deepseek v3

  #2803 closed Mar 5, 2025

• [Feature] When will pipeline model parallelism be supported?

  #4059 closed Mar 5, 2025

• [Bug] Config file not found when use NVIDIA_H20-3e

  #4028 closed Mar 5, 2025

• [willing to PR] optimzation of eagle2

  #2720 closed Mar 5, 2025

• [Bug] cannot build sgl-kernel without GPU

  #4060 closed Mar 4, 2025

• [Bug] Directly importing Grafana JSON does not work

  #4050 closed Mar 4, 2025

• [Bug] LoRA Model Not Appearing or Working with sglang Server

  #4057 closed Mar 4, 2025

• logger "Receive: obj=GenerateReqInput()" part with text rather than input_ids.

  #4045 closed Mar 4, 2025

• [Feature] Enable SGLang on more AMD GPUs

  #2320 closed Mar 4, 2025

• [Feature] Possible optimization in actor rollout parameter sync

  #2708 closed Mar 4, 2025

• [Feature] Make vLLM optional in model code

  #1673 closed Mar 3, 2025

• Development Roadmap (2025 H1)

  #4035 closed Mar 3, 2025

• Development Roadmap (2024 Q4)

  #1487 closed Mar 3, 2025

• [Feature] Split Docs CI

  #3158 closed Mar 3, 2025

• [Bug] Min new tokens error

  #4011 closed Mar 3, 2025

• [Bug] After using H200*8 to deploy DeepSeekR1, the large stress test model crashes

  #4019 closed Mar 3, 2025

• [Feature] Do we have any plan to integrate FLashMLA

  #4006 closed Mar 3, 2025

• [Bug] AWQ scalar type error

  #3780 closed Mar 3, 2025

• [Feature] Reward EOS close to max_tokens

  #2694 closed Mar 3, 2025

• [Bug] Continuous batching (OpenAI Server) with greedy search return different results

  #2687 closed Mar 3, 2025

• [Bug] loading phi4-mini-instruct with sglang

  #3935 closed Mar 2, 2025

• [Bug] After using --enable-torch-compile, garbled output

  #3944 closed Mar 2, 2025

• [Bug] H20 deepseek infer enable flashinfer mla hang

  #3578 closed Mar 2, 2025

52 Issues opened by 46 people

• [Bug] The sgl-kernel 0.0.3.post7 can't pass the CIs.

  #4214 opened Mar 8, 2025

• [Feature] sglang-router should perform extra status check on workers upon startup in addition to port reachability

  #4208 opened Mar 8, 2025

• [Bug] I encountered a 'Capture CUDA graph failed' error.

  #4205 opened Mar 8, 2025

• [Bug] After enabling flashinfer-mla for DeepSeek R1, I observed no throughput performance improvement.

  #4204 opened Mar 8, 2025

• [Bug] text generation hangs after serving some requests

  #4191 opened Mar 8, 2025

• [Bug] Server crashes with CUDA errors during EAGLE Speculative Decoding under high concurrency

  #4188 opened Mar 7, 2025

• Our performance test of DeepSeek-R1-Block-INT8 is inconsistent with #3730

  #4180 opened Mar 7, 2025

• [Feature] Support correctly exit using ctrl+c

  #4173 opened Mar 7, 2025

• [Bug] Qwen2.5-VL-7B-Instruct Inference Server crashes

  #4171 opened Mar 7, 2025

• [Bug] DeepSeek server crushed while using sglang.bench_serving

  #4161 opened Mar 7, 2025

• [Bug] sglang-router failure when first load model, try again successed

  #4160 opened Mar 7, 2025

• [Bug] Key conflict of `AutoImageProcessor.register`

  #4159 opened Mar 7, 2025

• [Bug] Accuracy issue with SGLang using DeepSeek-R1-AWQ

  #4158 opened Mar 7, 2025

• [Bug] --enable-metrics raise Error

  #4157 opened Mar 7, 2025

• [Feature] Support multiple Languages for Docs

  #4155 opened Mar 7, 2025

• [Bug] Eagle error with small vocal

  #4153 opened Mar 6, 2025

• [Bug] Incorrect DP rank and DP worker messaging when using TP > 1 and DP > 1

  #4149 opened Mar 6, 2025

• [Bug] Is YaRN supported in SGLang? How to enable it?

  #4145 opened Mar 6, 2025

• [Bug] HiRadixCache.__init__() got an unexpected keyword argument 'token_to_kv_pool_allocator'

  #4143 opened Mar 6, 2025

• [Bug] trying to build sglang and run on Jetson AGX orin: RuntimeError: RMSNorm failed with error code no kernel image is available for execution on the device

  #4130 opened Mar 6, 2025

• [Feature] nvcc fatal : Unknown option '-generate-dependencies-with-compile'

  #4120 opened Mar 6, 2025

• [Bug] NEXTN CUDA GRAPH master node use more memory (20GB) than worker node

  #4119 opened Mar 6, 2025

• [Bug] launch_server.py: error: unrecognized arguments: --enable-reasoning

  #4116 opened Mar 6, 2025

• [Bug] How to use SGLang's cpu code for inference

  #4098 opened Mar 5, 2025

• [Bug] -inf in top_logprobs breaks FastAPI endpoint

  #4097 opened Mar 5, 2025

• [Bug] Processes hang after weight update process group initialized

  #4096 opened Mar 5, 2025

• [Bug] AssertionError: res=<Response [503]> Process was always killed automactically

  #4094 opened Mar 5, 2025

• [Bug] NEXTN performance becomes low after some hours

  #4093 opened Mar 5, 2025

• [Bug] 'DeepseekV3ForCausalLM' object has no attribute 'get_embed_and_head'

  #4092 opened Mar 5, 2025

• [BUG] The #3836 problem still exists

  #4091 opened Mar 5, 2025

• Questions about the calculation of `max_req_num`

  #4090 opened Mar 5, 2025

• [Bug] RecursionError: maximum recursion depth exceeded while calling a Python object

  #4088 opened Mar 5, 2025

• [Bug] cuda kernel illegal and GPTQMarlinMoEMethod.apply() got an unexpected keyword argument 'correction_bias'

  #4083 opened Mar 5, 2025

• [Bug] Tool call with Llama3 models has inconsistent behavior with OpenAI

  #4072 opened Mar 4, 2025

• [Bug] granite-vision-3.2-2b failing on sglang with "LlavaNextForConditionalGeneration not supported"

  #4062 opened Mar 4, 2025

• [Feature] Add e4m3fnuz support to MoE-EP in FP8

  #4056 opened Mar 4, 2025

• [Feature] Apply structured output sampling after reasoning steps in Reasoning models

  #4055 opened Mar 4, 2025

• NVIDIA L40*8 docker NCCL Hanging During Initialization on Single Node with Multiple GPUs

  #4054 opened Mar 4, 2025

• [Error]Input length (160062 tokens) exceeds the maximum allowed length (59862 tokens).

  #4048 opened Mar 4, 2025

• [Bug] ImportError: cannot import name 'BaseImageProcessor' from 'transformers'

  #4047 opened Mar 4, 2025

• Development Roadmap (2025 H1)

  #4042 opened Mar 4, 2025

• [Feature] Refactor all parser features

  #4036 opened Mar 3, 2025

• [Bug] How to add chat_template on Offline Batch Inference

  #4024 opened Mar 3, 2025

• [Bug] running requests low

  #4022 opened Mar 3, 2025

• [Bug] After using H200*8 to deploy DeepSeekR1, the large stress test model crashes

  #4020 opened Mar 3, 2025

• [Bug] KeyError: 'model.layers.0.mlp.down_proj.weight_scale_inv' when run deepseek 671b with 64 RTX 4090 GPU

  #4018 opened Mar 3, 2025

• [Bug] deepseek-r1 occasionally crash

  #4017 opened Mar 3, 2025

• [Bug] Stuck at CUDA graph capture when serving with two A100*8 nodes

  #4007 opened Mar 3, 2025

• [Bug] The ncu command blocks when collecting data.

  #4005 opened Mar 3, 2025

• [Feature] i can not use function call of deepseek-v3、R1 with sglang==0.4.3.post2.

  #4004 opened Mar 3, 2025

• Performance issue comparing on MI210x4

  #3996 opened Mar 2, 2025

• [Bug] [DeepSeek-R1/V3] The description of --kv-cache-dtype in the documentation and the code is inconsistent.

  #3995 opened Mar 2, 2025

150 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• model: Support Janus-pro

  #3203 commented on Mar 8, 2025 • 22 new comments

• Minicpmo

  #3023 commented on Mar 8, 2025 • 4 new comments

• Add device detection and count functions to utils.

  #3962 commented on Mar 8, 2025 • 4 new comments

• [ROCm] add amd ocp fp8 E3M4FUZ constant

  #3959 commented on Mar 8, 2025 • 4 new comments

• Apply sgl w8a8 fp8 kernel

  #3148 commented on Mar 8, 2025 • 3 new comments

• [ROCm] Enable per token group quant fp8 in amd

  #3702 commented on Mar 8, 2025 • 2 new comments

• IPv6 support

  #3949 commented on Mar 8, 2025 • 2 new comments

• [Bug fixed] fixed the crash when enable the dp-attention on the single card

  #3958 commented on Mar 8, 2025 • 2 new comments

• Feat/support code completion

  #3612 commented on Mar 8, 2025 • 1 new comment

• fix: second_per_grid_ts should be used to get mrop position

  #3682 commented on Mar 8, 2025 • 1 new comment

• fix: Fix deprecated max_tokens param in openai ChatCompletionRequest

  #3122 commented on Mar 8, 2025 • 1 new comment

• model: Intern vl 2.5

  #3351 commented on Mar 8, 2025 • 1 new comment

• [Bug Fix] Add partial rotary factor support for Phi-4 and upgrade to transformers v4.49.0

  #3984 commented on Mar 8, 2025 • 1 new comment

• [Frontend] Support optional auto truncation

  #3721 commented on Mar 8, 2025 • 0 new comments

• Support building Dockerfile and Dockerfile.rocm using local repo.

  #3734 commented on Mar 8, 2025 • 0 new comments

• add bench for OpenAI Chat Completions

  #3738 commented on Mar 8, 2025 • 0 new comments

• feat: log input text for OpenAI format API

  #3699 commented on Mar 8, 2025 • 0 new comments

• [enh] add '\n' for non-streaming case to keep connection alive

  #3689 commented on Mar 8, 2025 • 0 new comments

• split ReplicatedLinear used in MLA prefill computing along hidden_states[0] to save duplicated computing on all devices

  #3688 commented on Mar 8, 2025 • 0 new comments

• add switch to disable open api doc

  #3744 commented on Mar 8, 2025 • 0 new comments

• Make Marlin incompatible AWQ models work

  #3759 commented on Mar 8, 2025 • 0 new comments

• [BugFix] Illegal memory access for MoE On H20

  #3779 commented on Mar 8, 2025 • 0 new comments

• [moe] fix: correct the cache size in the last chunk

  #3679 commented on Mar 8, 2025 • 0 new comments

• [WIP][Feautre, Hardware] add initial suport for ascend npu

  #3782 commented on Mar 8, 2025 • 0 new comments

• [Fix]: DeepSeek support cpu_offload_gb option

  #3675 commented on Mar 8, 2025 • 0 new comments

• [ROCm] Enable MTP (NextN) on AMD GPU

  #3670 commented on Mar 8, 2025 • 0 new comments

• RPD integration into SGLANG with --profile-req support

  #3653 commented on Mar 8, 2025 • 0 new comments

• [docker] Distributed Serving with k8s Statefulset ( good example for DeepSeek-R1)

  #3631 commented on Mar 8, 2025 • 0 new comments

• Update decode kernel benchmark with new Triton backend interface

  #3618 commented on Mar 8, 2025 • 0 new comments

• solve the mrope position bugs for qwen2-vl

  #3605 commented on Mar 8, 2025 • 0 new comments

• [DRAFT] change block size&wave number to increase wave inflight

  #3565 commented on Mar 8, 2025 • 0 new comments

• Add README for benchmarking SGLang using GenAI-Perf.

  #3552 commented on Mar 8, 2025 • 0 new comments

• chore: use sglang as alias of sglang.launch_server

  #3546 commented on Mar 8, 2025 • 0 new comments

• Add integration with lws for multi-node inference

  #3540 commented on Mar 8, 2025 • 0 new comments

• Make decord package optional.

  #3528 commented on Mar 8, 2025 • 0 new comments

• [optimize][scheduler] implement shortest-queue method

  #3526 commented on Mar 8, 2025 • 0 new comments

• [ROCm] accelerate rocm image build

  #3481 commented on Mar 8, 2025 • 0 new comments

• fix adjust_max_prefix_ids bug

  #3478 commented on Mar 8, 2025 • 0 new comments

• Support MLA in Torch Native Attention Backend

  #3475 commented on Mar 8, 2025 • 0 new comments

• Check for Already Terminated Process before kill

  #3474 commented on Mar 8, 2025 • 0 new comments

• Support `n` in OpenAI API completions

  #3446 commented on Mar 8, 2025 • 0 new comments

• run eagle speculative decodeing error!

  #3362 commented on Mar 2, 2025 • 0 new comments

• HotFix: json serialization error when using OAI v1/batches endpoint with logprobs

  #3896 commented on Mar 8, 2025 • 0 new comments

• Support FP4 gemm (1/2)

  #3899 commented on Mar 8, 2025 • 0 new comments

• Add block_wise INT8 support MTP NextN function

  #3911 commented on Mar 8, 2025 • 0 new comments

• Docs; Fix link in offline enine

  #3919 commented on Mar 8, 2025 • 0 new comments

• [Feature] Support for qwen2vl radix cache

  #3924 commented on Mar 8, 2025 • 0 new comments

• [bugfix]fix finish reason in response

  #3926 commented on Mar 8, 2025 • 0 new comments

• [Feature] Support Efficient Sparse HiP Attention (InfiniteHiP) with Long-Context Generalization and KV Offloading Capabilties

  #3930 commented on Mar 8, 2025 • 0 new comments

• Update vLLM version for ROCm

  #3937 commented on Mar 8, 2025 • 0 new comments

• triton python backend

  #3939 commented on Mar 8, 2025 • 0 new comments

• [Feature] Devin/1740716491 Add a hash for each new release

  #3947 commented on Mar 8, 2025 • 0 new comments

• add deepgemm into sgl-kernel

  #3957 commented on Mar 8, 2025 • 0 new comments

• Update function_call_parser.py

  #3960 commented on Mar 8, 2025 • 0 new comments

• example: add async offline inference demo

  #3961 commented on Mar 8, 2025 • 0 new comments

• feat(remote_model): support variable remote backend for model loader

  #3964 commented on Mar 8, 2025 • 0 new comments

• Blackwell fp8 blockscale

  #3968 commented on Mar 8, 2025 • 0 new comments

• [CI] Remove unused imports with Ruff to pre-commit config

  #3969 commented on Mar 8, 2025 • 0 new comments

• FP4 weight loading and inference (2/2)

  #3972 commented on Mar 8, 2025 • 0 new comments

• Move `aiohttp` into public dependencies

  #3980 commented on Mar 8, 2025 • 0 new comments

• [ROCM MOE] Enable ROCM AITER Block MOE For DeepSeek R1/V3

  #3788 commented on Mar 8, 2025 • 0 new comments

• Workaround for transformers 4.49

  #3792 commented on Mar 8, 2025 • 0 new comments

• [Bug] Fix chat completion empty input

  #3793 commented on Mar 8, 2025 • 0 new comments

• Add model name in EntryClass for tracking doc update to date

  #3800 commented on Mar 8, 2025 • 0 new comments

• use torch api to get distributed backend

  #3804 commented on Mar 8, 2025 • 0 new comments

• Ensure Usage Data in Streaming Responses Aligns with vLLM’s Implementation

  #3814 commented on Mar 8, 2025 • 0 new comments

• [BugFix]: Add type check before compare for max_new_tokens

  #3832 commented on Mar 8, 2025 • 0 new comments

• Resolve LinearBase circular import

  #3833 commented on Mar 8, 2025 • 0 new comments

• [WIP] [Feature] Support DeepSeek-v3 gptq

  #3834 commented on Mar 8, 2025 • 0 new comments

• [Doc] Fix typo in backend/sampling_params

  #3835 commented on Mar 8, 2025 • 0 new comments

• [docs] Update outdated description about `torch.compile`

  #3844 commented on Mar 8, 2025 • 0 new comments

• Update FP8 kernel configuration for 4xGPU support on AMD

  #3850 commented on Mar 8, 2025 • 0 new comments

• [Feature] Support for Ascend NPU backend

  #3853 commented on Mar 8, 2025 • 0 new comments

• Fix block_shape in benchmark_vllm_vs_sglang_fused_moe_triton.py

  #3858 commented on Mar 8, 2025 • 0 new comments

• Feat: Supports returning expected error responses to wrong requests

  #3875 commented on Mar 8, 2025 • 0 new comments

• Update router.rs

  #3877 commented on Mar 8, 2025 • 0 new comments

• [Quantization] support VPTQ

  #3879 commented on Mar 8, 2025 • 0 new comments

• Add examples for token in token out for LLM

  #3886 commented on Mar 8, 2025 • 0 new comments

• [Feature] Prefill assistant response

  #3971 commented on Mar 5, 2025 • 0 new comments

• [Bug] RecursionError: maximum recursion depth exceeded while calling a Python object

  #3953 commented on Mar 5, 2025 • 0 new comments

• [Bug] [AMD] ncclAllReduce Error under tp > 1 + multi nodes + enable cuda graph when running DeepSeek R1 on MI300X

  #3622 commented on Mar 5, 2025 • 0 new comments

• [Bug] ImportError: cannot import name 'is_valid_list_of_images' from transformers.models.mllama.image_processing_mllama

  #3878 commented on Mar 5, 2025 • 0 new comments

• [Bug] Qwen 2.5 VL

  #3321 commented on Mar 5, 2025 • 0 new comments

• [Track] long context performance sglang vs vllm

  #3471 commented on Mar 6, 2025 • 0 new comments

• [Feature] Lora Development Roadmap

  #2929 commented on Mar 6, 2025 • 0 new comments

• [Bug] 8xMI300X DeepSeek v3/R1 takes 45 mins to download and 1 hour to load shards

  #3825 commented on Mar 6, 2025 • 0 new comments

• [Feature] Proposal for adding PD-Disaggregation Feature to SGLang

  #3554 commented on Mar 6, 2025 • 0 new comments

• [Bug] SGLang Tool Calling for Qwen2.5 models returns empty ChatCompletionMessage content

  #3797 commented on Mar 6, 2025 • 0 new comments

• [Bug] Parameter/message body error directly returns 500

  #3805 commented on Mar 7, 2025 • 0 new comments

• [Feature] remove vllm _custom_ops

  #2965 commented on Mar 7, 2025 • 0 new comments

• [Bug] fused_moe OOM when run deepseek-r1 with --speculative-algo NEXTN

  #3633 commented on Mar 7, 2025 • 0 new comments

• [Feature] Instructions for running Sglang on AMD RX 7900 XTX (gfx1100) ROCm 6.2.4

  #3243 commented on Mar 7, 2025 • 0 new comments

• [DeepseekR1]How ragged prefill manage kv_cache?

  #3849 commented on Mar 8, 2025 • 0 new comments

• [Bug] --dp-size issue with AMD 8xMI300X and Llama 3.1 70B

  #3890 commented on Mar 8, 2025 • 0 new comments

• Extensive benchmarking of reasoning models including variance

  #3725 commented on Mar 8, 2025 • 0 new comments

• [Feature] A800 kernel config file to support deepseek r1 bf16

  #3748 commented on Mar 9, 2025 • 0 new comments

• [Bug] GGUF tokenizations issues

  #3427 commented on Mar 2, 2025 • 0 new comments

• [Bug] HuggingFace and SGLang inference don't match

  #2671 commented on Mar 3, 2025 • 0 new comments

• [Bug] torch.distributed.all_reduce raised Segmentation fault on 2 * 8 * H800

  #3745 commented on Mar 3, 2025 • 0 new comments

• [Feature] Support Deepseek's DeepGemm MoE

  #3881 commented on Mar 3, 2025 • 0 new comments

• [Bug] Dimension mismatched error when capturing cuda graph while enabling NEXTN

  #3891 commented on Mar 3, 2025 • 0 new comments

• [Bug] tensor_model_parallel_all_reduce' is not defined

  #2931 commented on Mar 3, 2025 • 0 new comments

• [Bug] ERROR: No matching distribution found for vllm==0.6.3.post2.dev1; extra == "srt-hip"

  #3189 commented on Mar 3, 2025 • 0 new comments

• [Bug] AttributeError: module 'vllm._custom_ops' has no attribute 'silu_and_mul'

  #3392 commented on Mar 3, 2025 • 0 new comments

• [Bug] Decode Throughput Inconsistency Between bench_serving and Engine Logs

  #3050 commented on Mar 3, 2025 • 0 new comments

• [Feature] attention dp + attention tp for deepseek v3

  #3750 commented on Mar 3, 2025 • 0 new comments

• [Bug] KeyError: 'model.layers.0.self_attn.k_scale'

  #3936 commented on Mar 4, 2025 • 0 new comments

• [Track] progress in removing vLLM dependencies

  #2245 commented on Mar 4, 2025 • 0 new comments

• [Question] Why is the performance worse after --enable-flashinfer-mla on H20

  #3917 commented on Mar 4, 2025 • 0 new comments

• [Feature] Support mistralai/Pixtral

  #2351 commented on Mar 4, 2025 • 0 new comments

• [Bug] deploy the DeepSeek-R1-awq get jumbled or nonsensical answers

  #3580 commented on Mar 4, 2025 • 0 new comments

• The number of image token (3) should be the same as in the number of provided images (1)

  #3819 commented on Mar 4, 2025 • 0 new comments

• Question sgl_kernel on amd paltforms

  #3965 commented on Mar 4, 2025 • 0 new comments

• [Feature] Use xgrammar as default grammar backend to aviod I/O errors while using Outlines in a multi-node setting

  #3383 commented on Mar 5, 2025 • 0 new comments

• [WIP] Integration of TurboMind AWQ

  #2900 commented on Mar 8, 2025 • 0 new comments

• [Core] Optimize the delay scheduling of in batch prefix caching

  #2962 commented on Mar 8, 2025 • 0 new comments

• Integrate turbomind into sgl-kernel

  #2999 commented on Mar 8, 2025 • 0 new comments

• support telechat2 model

  #3000 commented on Mar 8, 2025 • 0 new comments

• Support int8 kvcahe

  #3034 commented on Mar 8, 2025 • 0 new comments

• Modify the kernel test path & add it to the CI process.

  #3044 commented on Mar 8, 2025 • 0 new comments

• [Feature] Beam Search

  #3066 commented on Mar 8, 2025 • 0 new comments

• [Feature] Rewrite Sampling Parameter #3165

  #3185 commented on Mar 8, 2025 • 0 new comments

• Add logit bias into the SGLang interface.

  #3187 commented on Mar 8, 2025 • 0 new comments

• Add deepseek_v3 fused gate

  #3191 commented on Mar 8, 2025 • 0 new comments

• add date_string to the chat template

  #3297 commented on Mar 8, 2025 • 0 new comments

• FEA compat with ipv6

  #3301 commented on Mar 8, 2025 • 0 new comments

• fix(mig): fallback gpu_memory_total value

  #3353 commented on Mar 8, 2025 • 0 new comments

• Better support of tp checkpoint loading

  #3367 commented on Mar 8, 2025 • 0 new comments

• Update Intel XPU install instruction

  #3390 commented on Mar 8, 2025 • 0 new comments

• fix: return null instead of "" for finish_reason when unfinished

  #3391 commented on Mar 8, 2025 • 0 new comments

• Modify metrics service endpoint

  #3443 commented on Mar 8, 2025 • 0 new comments

• Add torch profiler activity for HPU

  #3445 commented on Mar 8, 2025 • 0 new comments

• How to return reasoning_content from sglang server response?

  #3428 commented on Mar 9, 2025 • 0 new comments

• DeepSeek-R1 Optimization Option Ablations

  #3956 commented on Mar 9, 2025 • 0 new comments

• [Feature] Add initial support for sequence parallelism

  #1436 commented on Mar 8, 2025 • 0 new comments

• Surpport kv cache int8/int4 for triton backend

  #1644 commented on Mar 8, 2025 • 0 new comments

• feat: use cascade attention kernel (single level)

  #2101 commented on Mar 8, 2025 • 0 new comments

• Add support for Phi3V

  #2383 commented on Mar 8, 2025 • 0 new comments

• Add InfiniteBench for long context benchmarking

  #2421 commented on Mar 8, 2025 • 0 new comments

• [Experimental] Add a gRPC server for completion request

  #2478 commented on Mar 8, 2025 • 0 new comments

• Refactor Scheduler to improve code organization

  #2593 commented on Mar 8, 2025 • 0 new comments

• [Feature] support compute-communication overlap with TransformerEngine

  #2627 commented on Mar 8, 2025 • 0 new comments

• Support InternVL2 Series

  #2629 commented on Mar 8, 2025 • 0 new comments

• [Feature] Support regex as a stopping condition

  #2699 commented on Mar 8, 2025 • 0 new comments

• Speculative decoding with lookahead

  #2790 commented on Mar 8, 2025 • 0 new comments

• Add endpoint for file support, purely to speed up processing of input_embeds.

  #2797 commented on Mar 8, 2025 • 0 new comments

• [Feature] Support Deepseek-VL2

  #2798 commented on Mar 8, 2025 • 0 new comments

• Improve the mixed chunk prefill by lanuch two kernels

  #2811 commented on Mar 8, 2025 • 0 new comments

• Use CUDA_VISIBLE_DEVICES instead of gpu_id variables everywhere.

  #2824 commented on Mar 8, 2025 • 0 new comments

• [Feature] Support dynamic loading and unloading of Lora adapters

  #2891 commented on Mar 8, 2025 • 0 new comments