Support server based rollout in Verlengine #4848
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
13 tasks
|
Great work Chengxing, I will review all the design docs, dev logs and codes carefully. |
zhaochenyang20
requested review from
merrymercy,
Ying1123,
zhyncs,
hnyls2002,
ispobock and
ByronHsu
as code owners
|
I'm trying out this PR but got the error below: |
Chenxing, could you try to help on this? |
Sure! |
|
@yangky11 However, in your current implementation, you're calling The correct order of operations should be as follows: def update_weights_from_tensor(
self,
named_tensors: List[Tuple[str, torch.Tensor]],
load_format: Optional[str] = None,
):
# Most naive implementation, can optimize a lot if it is bottleneck
for tensor_index, (name, tensor) in enumerate(named_tensors):
serialized_tensor = MultiprocessingSerializer.serialize(
_preprocess_tensor_for_update_weights(tensor)
)
if self._tp_rank == 0:
gathered_serialized_tensors = [None for _ in range(self._tp_size)]
else:
gathered_serialized_tensors = None
dist.gather_object(
obj=serialized_tensor,
object_gather_list=gathered_serialized_tensors,
dst=self._device_mesh_cpu.mesh.tolist()[0],
group=self._device_mesh_cpu.get_group(),
)
if self._tp_rank == 0:
self._engine.update_weights_from_tensor(
named_tensors=[
(
name,
LocalSerializedTensor(values=gathered_serialized_tensors),
)
],
load_format=load_format,
flush_cache=tensor_index == len(named_tensors) - 1,
)and then in def update_weights_from_tensor(
self,
named_tensors: List[Tuple[str, torch.Tensor]],
load_format: Optional[str] = None,
flush_cache: bool = False,
):
print(f"update_weights_from_tensor of HttpServerEngineAdapter")
return requests.post(
f"http://localhost:{self.server_args.port}/update_weights_from_tensor",
json={
"serialized_named_tensors": [
serialize_for_http(
MultiprocessingSerializer.serialize(named_tensors)
)
for _ in range(self.server_args.tp_size)
],
"load_format": load_format,
"flush_cache": flush_cache,
},
) |
|
@yitianlian That makes sense! I'm trying to have a minimal example that performs the preprocessing in Like before, I launched the SGLang server using import requests
import pickle
import base64
from transformers import AutoModel
from torch.distributed.tensor import DTensor
from sglang.srt.utils import MultiprocessingSerializer
from sglang.srt.model_executor.model_runner import LocalSerializedTensor
def serialize_for_http(data):
# First pickle the data, then convert to base64 for safe HTTP transmission
pickled = pickle.dumps(data)
return base64.b64encode(pickled).decode("utf-8")
model = AutoModel.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
name, tensor = list(model.state_dict().items())[0]
print(name, tensor)
serialized_tensor = MultiprocessingSerializer.serialize(
tensor.full_tensor() if isinstance(tensor, DTensor) else tensor
)
print(serialized_tensor)
gathered_serialized_tensors = [serialized_tensor]
requests.post(
"http://127.0.0.1:30000/update_weights_from_tensor",
json={
"serialized_named_tensors": [
serialize_for_http(
MultiprocessingSerializer.serialize(
[(name, LocalSerializedTensor(values=gathered_serialized_tensors))]
)
)
],
"load_format": None,
"flush_cache": False,
},
)Here is the error I got: I'm wondering if I'm still misunderstanding how |
|
cc @fzyzcjy . Hi Tom, would you mind have a check on this inter-process communication issue? Huge thanks! |
|
@yangky11 Hi, the implementation approach in this PR (HttpServerEngineAdapter, update_weights_from_tensor in http server, etc) was proposed by me, thus I will handle this. The digest issue seems to because that, a tensor on CPU is being sent across processes (when calling update weights) and the two processes do not have parent-child relation. Therefore, make the tensor on GPU will solve the problem. If the tensor needs be on CPU for some reasons (would like to know a bit why, since it seems to introduce extra CPU-GPU copies in RL scenario), I can make a PR about it. Btw I am also doing AI for mathematics (as an individual researcher), and am interested in engineering such as SGLang. If there are things that SGLang needs to do for AI4math, I am happy to discuss. |
fzyzcjy
reviewed
Apr 4, 2025
Comment on lines
416
to
417
| import base64 | ||
| import pickle |
There was a problem hiding this comment.
nit: maybe we can move this import to top of file
Comment on lines
424
to
430
| @app.post("/update_weights_from_tensor") | ||
| async def update_weights_from_tensor( | ||
| obj: UpdateWeightsFromTensorReqInput, request: Request | ||
| ): | ||
| obj.serialized_named_tensors = [ | ||
| deserialize_from_http(item) for item in obj.serialized_named_tensors | ||
| ] |
There was a problem hiding this comment.
wondering whether we should make the API like, Union[UpdateWeightsFromTensorReqInput, str] (and the "str" means serializing the whole requests into a string using base64), because this may be a bit more general
Comment on lines
17
to
20
| from sglang.test.test_utils import ( | ||
| DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH, | ||
| popen_launch_server, | ||
| ) |
There was a problem hiding this comment.
nit: may not be very ideal if we import test code from main code
Comment on lines
23
to
26
| def serialize_for_http(data): | ||
| # First pickle the data, then convert to base64 for safe HTTP transmission | ||
| pickled = pickle.dumps(data) | ||
| return base64.b64encode(pickled).decode("utf-8") |
There was a problem hiding this comment.
nit: maybe we can put serialize_for_http and deserialize_from_http, into one common place, because they are pairs that will be used together (for example, maybe write like MultiprocessingSerializer)
| return base64.b64encode(pickled).decode("utf-8") | ||
|
|
||
|
|
||
| import dataclasses |
There was a problem hiding this comment.
nit: would be great to make import at top of file
| kwargs["log_level"] = "error" | ||
| server_args = ServerArgs(**kwargs) | ||
| if self._tp_rank == 0: | ||
| self._engine = HttpServerEngineAdapter(server_args) |
There was a problem hiding this comment.
wondering whether we are missing tp_size, node_rank, nnodes here
| kwargs["log_level"] = "error" | ||
| server_args = ServerArgs(**kwargs) | ||
| if self._tp_rank == 0: | ||
| self._engine = HttpServerEngineAdapter(server_args) |
There was a problem hiding this comment.
to make Engine and HttpServerEngineAdapter be as identical as possible, it would be great to let HttpServerEngineAdapter accept same args as Engine, i.e. accept kwargs instead of ServerArgs.
Comment on lines
51
to
59
| if "server_args" in kwargs: | ||
| # Directly load server_args | ||
| server_args = kwargs["server_args"] | ||
| else: | ||
| # Construct server_args from kwargs | ||
| if "log_level" not in kwargs: | ||
| # Do not print logs by default | ||
| kwargs["log_level"] = "error" | ||
| server_args = ServerArgs(**kwargs) |
There was a problem hiding this comment.
Hmm looks like the code is copied from Engine, if so then
- it would be great to avoid copying code, e.g. extract a function
compute_server_args_from_kwargs - the layering is a little bit weird if we keep it: the same logic, "computing serverargs from kwargs", is put (a) in Engine/Adapter layer, and (b) in the VerlEngine layer which is one layer above Engine/Adapter. thus would be great to move it to Adapter.
| os.environ["SGLANG_BLOCK_NONZERO_RANK_CHILDREN"] = "0" | ||
| self._engine = Engine( | ||
| **kwargs, tp_size=self._tp_size, node_rank=node_rank, nnodes=nnodes | ||
| ) | ||
| elif "launch_server" in kwargs and kwargs["launch_server"]: |
There was a problem hiding this comment.
the top-level if-elif-else currently does:
- if: for Engine, part 1
- elif: for Adapter
- else: for Engine, part 2
Thus would be great to have
if backend == 'engine':
if first rank in node: self._engine = Engine()
else: self._engine = None
elif backend == 'server':
if rank zero: self._engine = Adapter()
else: self._engine = None
else:
raise error
There was a problem hiding this comment.
given that "VerlEngine using Engine" and "VerlEngine using Adapter" should behave almost identical to users, maybe we should not create a new test, but instead change some lines to the original test. This can both reduce code duplication, and also ensure we are (implicitly) checking they behave similarly.
(I have not checked this file - will check after this change)
|
Given the comments from @fzyzcjy , I've rewritten my code and addressed most of the issues that were raised. However, there are still a few points that need further discussion:
|
yangky11
reviewed
Apr 4, 2025
| HttpSerializer.serialize( | ||
| MultiprocessingSerializer.serialize(named_tensors) | ||
| ) | ||
| for _ in range(self.server_args.tp_size) |
There was a problem hiding this comment.
I'm wondering if here we can call the serialization only once.
There was a problem hiding this comment.
I remember MultiprocessingSerializer.serialize(named_tensors) can't be posted by HTTP?
There was a problem hiding this comment.
I meant something like
x = HttpSerializer.serialize(MultiprocessingSerializer.serialize(named_tensors))
response = requests.post(self._url("update_weights_from_tensor"), json={"serialized_named_tensors": [x for _ in range(self.server_args.tp_size)] ...
There was a problem hiding this comment.
Oh, sure! I will fix it now.
There was a problem hiding this comment.
Or we can just send a single copy of HttpSerializer.serialize(MultiprocessingSerializer.serialize(named_tensors)) to reduce the HTTP payload size? The server can make multiple copies after receiving it.
There was a problem hiding this comment.
The size of serialized tensors is really small, so I think it will not take a long time. Also, it seems that the update_weight_from_tensor entry point doesn't have access to the tp size.
|
One way may be with Remarks
|
fzyzcjy
reviewed
Apr 10, 2025
| class HttpServerEngineForRL(EngineBase): | ||
| def __init__(self, **kwargs): | ||
| self.server_args = ServerArgs(**kwargs) | ||
| print(f"launch_server_from_verl_engine {self.server_args.port}") |
There was a problem hiding this comment.
yes (or at least rename the word "launch_server_from_verl_engine" which seems to be the old name of a function that we called
| print(f"launch_server_from_verl_engine {self.server_args.port}") | ||
| self.process = launch_server_process(self.server_args) | ||
|
|
||
| def _make_request(self, endpoint: str, payload: dict = None): |
There was a problem hiding this comment.
Suggested change
| def _make_request(self, endpoint: str, payload: dict = None): | |
| def _make_request(self, endpoint: str, payload: Optional[dict] = None): |
| flush_cache: bool = False, | ||
| ): | ||
| """ | ||
| Update model weights from tensor data. The HTTPS server will only post meta data, and the real weights will be copied directly from GPUs. |
There was a problem hiding this comment.
nit: it seems most people will use HTTP instead of HTTPS (indeed wondering whether SGLang supports https today), thus would be great to change doc
(same for other "HTTPS" words)
| - No pickle serialization is used for security reasons | ||
| """ | ||
|
|
||
| serialized_named_tensors: List[str] |
There was a problem hiding this comment.
Suggested change
| serialized_named_tensors: List[str] | |
| serialized_named_tensors: List[Union[str, bytes]] |
|
|
||
| - Binary data like tensors are base64 encoded | ||
| - Data is structured in JSON for easy transmission over HTTP | ||
| - No pickle serialization is used for security reasons |
There was a problem hiding this comment.
well there is pickle indeed... (to serialize torch.Tensors)
| flush_cache: bool | ||
| """Update model weights from tensor input. | ||
|
|
||
| - Binary data like tensors are base64 encoded |
There was a problem hiding this comment.
it is not base64 encoded when this object is created from Engine...
|
cc @fzyzcjy . Shout out to your great help and guidance! I quickly made the changes as you reviewed and suggested. Please let me know if there is anything left that we need to take care and discuss with! |
|
If the PR is in emergency then I have no big issues about it. |
zhaochenyang20
approved these changes
Apr 12, 2025
jimoosciuc
pushed a commit
to Furion-cn/sglang
that referenced
this pull request
Apr 17, 2025
Co-authored-by: Jin Pan <[email protected]> Co-authored-by: Chayenne <[email protected]> Co-authored-by: Jinn <[email protected]>
pi314ever
pushed a commit
to pi314ever/sglang
that referenced
this pull request
Apr 23, 2025
* Support with_stack and record_shapes in profiler (sgl-project#4740) Co-authored-by: Lianmin Zheng <[email protected]> * test: reduce `mem_fraction_static` for gemma3 vision test (sgl-project#4840) * Fix CI tests (sgl-project#4853) * Fix fa3 cuda graph page_size > 1 precision and page_size=1 speed (sgl-project#4855) * Revert "get the python version from env (sgl-project#4729)" (sgl-project#4863) * [Feature] add multi-rank support for Lora (sgl-project#4492) Co-authored-by: rudy152 <[email protected]> * Clean up `import vllm` in quantization/__init__.py (sgl-project#4834) * Fix wrong variable name when stopping memory profile (sgl-project#4772) * [Feat] support deepgemm for cmake (sgl-project#4864) * Make torch compile configurable for biased_grouped_topk (sgl-project#4749) * update sgl-kernel test ci (sgl-project#4866) * fix sampling issue (sgl-project#4871) * bump sgl-kernel 0.0.5.post4 (sgl-project#4768) * fix sgl-kernel cu118 build (sgl-project#4872) * [Feature] Support FA3 backend for MLA (sgl-project#4831) * upgrade sgl-kernel 0.0.5.post4 (sgl-project#4873) * update torch compile doc (sgl-project#4874) * bump v0.4.4.post3 (sgl-project#4878) * Fix BadRequestError wrong arguments and remove openai dependency (sgl-project#4882) * Improve stack trace of retry errors (sgl-project#4845) * Tiny fix doc error (sgl-project#4795) * [Docs] Update DeepGEMM at README.md (sgl-project#4886) * Update CODEOWNERS (sgl-project#4889) * Delete test_deep_gemm.py (sgl-project#4891) * Add deepseek style fused moe group gate selection kernel (sgl-project#4530) * quick fix: add default for new kernel (sgl-project#4898) * remove setup for sgl-kernel (sgl-project#4899) * [Misc] Clean m.def and add Development Tips (sgl-project#4890) * fix allreduce test (sgl-project#4909) * Support page size > 1 + eagle (sgl-project#4908) * Fix retract for page size > 1 (sgl-project#4914) * [Feature] use pytest for sgl-kernel (sgl-project#4896) * fix bmm fp8 (sgl-project#4926) * Fix the timeout for unit-test-2-gpu in pr-test.yml (sgl-project#4927) * Fix 2-gpu CI test and suppress some warnings (sgl-project#4930) * [feat] add fa3 in sgl-kernel (sgl-project#4902) Co-authored-by: Sleepcoo <[email protected]> * Fix sglang frontend's incorrect dependency on torch (sgl-project#4931) * [Fix] avoid stream sync and torch compile in prefill for fa3 backend (sgl-project#4932) * cleanup sgl-kernel (sgl-project#4933) * [Fix] Improve Lora tests and reduce CI runtime (sgl-project#4925) * Fix DeepSeek bug causing 2.2% MMLU drop when TP!=DP (sgl-project#4883) Co-authored-by: ch-wan <[email protected]> * [Fix] Add torch compile for torch.clamp back (sgl-project#4936) * Fix oom error for large page size (sgl-project#4913) Co-authored-by: Lianmin Zheng <[email protected]> * [feat] interface for platforms abstraction (sgl-project#4928) * [Fix] revert clean m.def for cudagraph (sgl-project#4944) * refactor: multimodal data (sgl-project#4754) * bump sgl-kernel v0.0.6 (sgl-project#4950) * [Build] Fix cuda12.8 build error in nvfp4_scaled_mm_kernels.cu (sgl-project#4953) * use fa3 in sgl-kernel (sgl-project#4954) * Revert PR 4764 & 4813 related to R1 RoPE (sgl-project#4959) * [Feature] Support DeepEP Low Latency (sgl-project#4767) Co-authored-by: sleepcoo <[email protected]> Co-authored-by: laixinn <[email protected]> Co-authored-by: ch-wan <[email protected]> * update bench_serving (sgl-project#4958) * Prevent memory leak of retract_decode when page_size > 1 (sgl-project#4977) * [VLM RLHF] Take Image input for verl vlm rollout (sgl-project#4915) Signed-off-by: Xinyuan Tong <[email protected]> Co-authored-by: GeLee <[email protected]> * Large page size aligned hierarchical caching (sgl-project#4581) * bug fix for hicache host eviction (sgl-project#4989) * sgl scaled_fp8_quant support output padding (sgl-project#4861) * Add Eagle Speculative Decoding to FA3 Backend (sgl-project#4951) Co-authored-by: hebiao064 <[email protected]> Co-authored-by: Baizhou Zhang <[email protected]> Co-authored-by: zcnrex <[email protected]> * Update tokenizer_manager.py (sgl-project#5008) * [sgl-kernel] per token group quant support COLUMN MAJOR (sgl-project#4817) * update cutlass tag (sgl-project#5011) * Feature/revise docs ci (sgl-project#5009) * fix: fix illegal cuda memory access at fused_moe_kernel (sgl-project#4727) Co-authored-by: yuethe <[email protected]> * [Build] Support build sgl-kernel with ccache (sgl-project#5020) * fix deepgemm as well (sgl-project#5030) * try to fix ci oserror (sgl-project#5024) * Replace enable_flashinfer_mla argument with attention_backend (sgl-project#5005) * Small refactor DeepEPMode to clean up code a bit (sgl-project#4992) * [Fix] fix fa3 build at cu118 (sgl-project#5036) * Revert "Replace enable_flashinfer_mla argument with attention_backend" (sgl-project#5048) * bump sgl-kernel v0.0.7 (sgl-project#5046) * update eagle-3 docs (sgl-project#4796) Co-authored-by: Yifan Zhang <[email protected]> * Add LlavaLlamaForCausaLM in MultiModal Processors (sgl-project#5039) Co-authored-by: Ravi Theja Desetty <[email protected]> * Update the retry count (sgl-project#5051) * upgrade sgl-kernel v0.0.7 (sgl-project#5049) * [2/3] fix dsv3 awq issue (sgl-project#4625) Co-authored-by: 晟海 <[email protected]> Co-authored-by: laixinn <[email protected]> * Feature/revise docs ci (sgl-project#5056) * Add H20 fused MoE kernel tuning configs for DeepSeek V3/R1 (sgl-project#5057) * [fix] remove `cuda_device_count_stateless` (sgl-project#5060) * Small refactor DeepEPDispatcher into subclasses (sgl-project#4994) * Support async DeepEP by splitting into two stages (sgl-project#4995) * Cleanup unused resources after DeepEP operation (sgl-project#4996) * Add DeepSeek V3/R1 shared experts fusion (sgl-project#4918) * [deepep] fix: shared experts are not initialized when shared experts fusion is enabled (sgl-project#5072) * fix dummy-load deepseekv2 (sgl-project#4535) * support sgl-kernel on blackwell (sgl-project#5074) * FA3 Spec Decoding to support top k = 1 and add cuda graph support (sgl-project#5050) Co-authored-by: Qingquan Song <[email protected]> Co-authored-by: Chunan Zeng <[email protected]> * [Revision] Replace enable_flashinfer_mla argument with attention_backend (sgl-project#5052) * upgrade transformers 4.51.0 (sgl-project#5088) * sgl-kernel transfer custom allreduce from trt kernel to vllm kernel (sgl-project#5079) * bump sgl-kernel 0.0.8 (sgl-project#5089) * python transfer custom allreduce from trt kernel to vllm kernel (sgl-project#5080) * bump v0.4.4.post4 (sgl-project#5091) * Fix: Reduce the number of document ci attempts to avoid long ci running (sgl-project#5097) Co-authored-by: shuaills <[email protected]> * Add Llama4 support (sgl-project#5092) Co-authored-by: Cheng Wan <[email protected]> Co-authored-by: fzyzcjy <[email protected]> Co-authored-by: ispobock <[email protected]> * Fix refactor error - fp8.py (sgl-project#5106) Co-authored-by: Lianmin Zheng <[email protected]> * bump v0.4.5 (sgl-project#5117) * [ci] fix llama4 ci error (sgl-project#5126) * Refactor and Optimize FA3 Code (sgl-project#5090) Co-authored-by: Qingquan Song <[email protected]> * Add Llama4 user guide (sgl-project#5133) Co-authored-by: Cheng Wan <[email protected]> * [Misc] Use pytest.mark.skipif in sgl-kernel test (sgl-project#5137) * feat: disable grammar restrictions within reasoning sections (sgl-project#4984) Co-authored-by: tianhaoyu <[email protected]> Co-authored-by: DarkSharpness <[email protected]> * [modelopt] automatically inspect if model is ModelOpt quantized and set quantization method (sgl-project#5145) * [AMD] Fix missing per_token_group_quant_fp8 for ROCm (sgl-project#5140) * fix multimodal hash feature (sgl-project#5083) * Fix run time error in ROCm platform (sgl-project#5147) Co-authored-by: wunhuang <[email protected]> Co-authored-by: root <[email protected]> * [FA3 Feature] Support multi modal Llama-3.2-11B-Vision-Instruct (sgl-project#5103) * Add unit test on page_size > 1 and mla and integration test for Flash Attention 3 (sgl-project#4760) * Use public model for FA3 speculative decode testing (sgl-project#5152) * Add dummy grok test to amd CI. (sgl-project#5115) * fix empty_cache error in pt_weights_iterator (sgl-project#5151) Co-authored-by: dangkai.dk <[email protected]> * Fix torch compile errors (sgl-project#5158) * Fix loading KV quantization scale; Enable modelopt kv cache (sgl-project#4686) Co-authored-by: qingquansong <[email protected]> * [PD] Fix unclosed prefill connection warning of mini_lb (sgl-project#5155) Signed-off-by: Shangming Cai <[email protected]> * Add optimized native kernels in sgl-kernel (sgl-project#5150) Co-authored-by: Chunyuan WU <[email protected]> Co-authored-by: YanbingJiang <[email protected]> Co-authored-by: blzheng <[email protected]> * [PD] Simplify mini LB (sgl-project#4911) Co-authored-by: Liangsheng Yin <[email protected]> * Small improvement of native api docs (sgl-project#5139) Co-authored-by: zhaochenyang20 <[email protected]> * [feat&refactor] Enhance multimodal input support with refactor io_struct (sgl-project#4938) Signed-off-by: Xinyuan Tong <[email protected]> * Support 2x8xH100 for Llama 4 (sgl-project#5159) * FP4 weight loading and inference (2/2) (sgl-project#3972) * Fix multimodal hashing error (sgl-project#5174) * Tiny disable model that does not work (sgl-project#5175) * [Bugfix] Fix index out of bounds in local attention with large sequences (sgl-project#5173) * [Fix] DeepEP Compatibility with Low Latency (sgl-project#5068) Co-authored-by: ch-wan <[email protected]> * docs: remove the use of Downward API for LWS_WORKER_INDEX (sgl-project#5110) Signed-off-by: Kay Yan <[email protected]> * feat: add DeepGEMM build warning (sgl-project#5176) Co-authored-by: grimoire <[email protected]> * fix: use DeepEPDispatcher on CUDA (sgl-project#5180) * [DeepEP] fix: import buffer error (sgl-project#5179) * Let `bench_one_batch` support `enable_dp_attention` (sgl-project#4058) * [Misc] clean up vllm in sgl-kernel test (sgl-project#5189) * Fix ci test "test_eval_fp8_accuracy" failed (sgl-project#5185) Co-authored-by: wunhuang <[email protected]> * Optimize topk operation in llama4 (sgl-project#5128) * Support Llama4 fp8 inference (sgl-project#5194) Co-authored-by: laixinn <[email protected]> Co-authored-by: sleepcoo <[email protected]> Co-authored-by: zhyncs <[email protected]> * [ci] fix ci test fused_moe op (sgl-project#5102) * model: support mllama4 (sgl-project#5144) * update grok test (sgl-project#5171) * sgl-kernel use cutlass latest version for fp8 blockwise gemm (sgl-project#5207) * Add H20 dtype fp8_w8a8 fused MoE kernel tuning configs for DeepSeek V3/R1 (sgl-project#5196) * fix: log warning when disable cuda graph (sgl-project#5209) * [metrics] Add in queue metrics (sgl-project#4444) * Fix DeepSeek error when using DeepEP mode (sgl-project#5190) * reduce moe_align_block_size_kernel small batch mode overhead (sgl-project#5086) * [PD] Support KV transfer with mooncake (sgl-project#4880) Signed-off-by: Shangming Cai <[email protected]> Co-authored-by: Shangming Cai <[email protected]> Co-authored-by: Xuchun Shang <[email protected]> Co-authored-by: shangmingc <[email protected]> * [PD] Add get_contiguous_buf_infos interface for MLATokenToKVPool (sgl-project#5204) * Update deps for mllama4 (sgl-project#5215) * Fix deepseek-v3 with torch.compile in PyTorch 2.6. (sgl-project#5213) * ROCm sgl-kernel: compatible to later torch (sgl-project#5167) * [Misc] Clean sgl-kernel test (sgl-project#5216) * Update Makefile / build script to avoid installing incompatible torch dependency (sgl-project#5245) * Fix torch.compile cacheing (sgl-project#5259) Co-authored-by: zhyncs <[email protected]> * ROCm/AITER CK_MoE: update 2-stage kernels & support both Activations (sgl-project#5228) * Optimize attention in llama4 (sgl-project#5127) * Optimize GPU memory usage in FlashAttentionBackend's strided indexing (sgl-project#5262) Co-authored-by: ch-wan <[email protected]> * Support `--enable-llama4-multimodal` (sgl-project#5254) * [fix] fix mrope positions not picked up (sgl-project#5265) * doc: nested loop code for offline engine (sgl-project#5244) * fix: examples for token_in_token_out_vlm (sgl-project#5193) * Fix a 404 link in send_request.ipynb (sgl-project#5280) Signed-off-by: windsonsea <[email protected]> * fix: enable fp4 compilation on cu128 (sgl-project#5286) * feat: add cu128 identifier for sgl-kernel (sgl-project#5287) * chore: relax the torch version restriction for sgl-kernel compilation (sgl-project#5288) * chore: bump sgl-kernel v0.0.8.post1 (sgl-project#5289) * [PD] fix: skip warmup request in disaggregation mode to prevent crash on timeout (sgl-project#5292) * [Docs] Supported Model Docs - Major restructuring (sgl-project#5290) Co-authored-by: zhaochenyang20 <[email protected]> * fix: update update_wheel_index for cu128 (sgl-project#5300) * [Docs] Remove the older supported docs section (sgl-project#5301) * remove moe_align_block_size torch.zeros in small batch/expert mode (sgl-project#5298) * feat: add blackwell Dockerfile (sgl-project#5302) * feat: add blackwell workflow (sgl-project#5303) * fix: use fa3 unit test on hopper only (sgl-project#5304) * misc: update blackwell Dockerfile (sgl-project#5306) * fix: remove cublas_grouped_gemm (sgl-project#5307) * fix: update flash attn (sgl-project#5308) * fix: use deepgemm only on hopper (sgl-project#5310) * [VLM] Adopt fast image processor by default (sgl-project#5065) * Adjust ci test threshold (sgl-project#5271) * Blackwell Cutlass MLA kernel (sgl-project#5142) * misc: cleanup 3rdparty (sgl-project#5311) * update variable naming and comments for rocm (sgl-project#5299) * Fix w8a8_int8 model shared experts fusion load weights error (sgl-project#5120) * Add flash_attn_varlen_func to sgl-kernel (sgl-project#5315) * Fix fa3 window size setup (sgl-project#5316) * chore: bump sgl-kernel v0.0.8.post2 (sgl-project#5317) * feat: use fa3 mla by default on hopper (sgl-project#5210) Co-authored-by: yundai424 <[email protected]> Co-authored-by: hebiao064 <[email protected]> * Fix: docs/backend/structured_outputs.ipynb (sgl-project#4884) * Delete python/sglang/srt/layers/moe/fused_moe_triton/configs/E=257,N=… (sgl-project#5321) * refine fused_moe tuning docs (sgl-project#5294) * Support server based rollout in Verlengine (sgl-project#4848) Co-authored-by: Jin Pan <[email protected]> Co-authored-by: Chayenne <[email protected]> Co-authored-by: Jinn <[email protected]> * [Feat] Add sparse attn to sgl-kernel (sgl-project#5327) * fix: solve cu118 issue for cutlass mla (sgl-project#5331) * chore: bump sgl-kernel v0.0.8.post3 (sgl-project#5332) * ci: update release node (sgl-project#5333) * fix: determine if flashinfer is installed (sgl-project#5336) * feat: adapt merge_state (sgl-project#5337) * misc: update sagemaker Dockerfile (sgl-project#5341) * Fix: Ensure tensors for dist.broadcast match NCCL backend device (sgl-project#5322) * docs: update adoption and sponsorship list with Oracle (sgl-project#5343) * chore: upgrade sgl-kernel 0.0.8.post3 (sgl-project#5342) * Fix typo: infight -> inflight (sgl-project#5357) * [PD] Add transfer backend abstraction (sgl-project#5328) * fix MLATokenToKVPoolHost get_size_per_token bug (sgl-project#5161) Co-authored-by: AniZpZ <[email protected]> * fix sgl-project#5322 (sgl-project#5359) * feat: update experiment_runner (sgl-project#5360) * [DeepEP] Reduce routed scaling overhead (sgl-project#5277) Co-authored-by: Cheng Wan <[email protected]> * Free metadata_buffer_index after transfer finished (sgl-project#5364) * Free metadata_buffer_index after transfer finished (sgl-project#5364) * Fix DeepSeek DP Attention + torch compile (sgl-project#5367) Co-authored-by: ispobock <[email protected]> * Support for Qwen2.5-VL Model in bitsandbytes Format (sgl-project#5003) * Fix PD disaggregation bugs (sgl-project#5326) * [PD Bug] fix MLA get_contiguous_buf_infos error (sgl-project#5384) * [perf] experimental enhance fp8 per-tensor quant (sgl-project#5370) * Apply deepseek cuda rope (sgl-project#5385) Co-authored-by: Yineng Zhang <[email protected]> * apply fused moe gate in ds v3/r1 (sgl-project#5371) Co-authored-by: Yineng Zhang <[email protected]> * fix: update test config (sgl-project#5392) * [Fix] Turn off DeepGEMM by default (sgl-project#5263) * minor clean up of sgl-kernel/CMakeLists.txt (sgl-project#5393) * Add A800 shared experts fused MoE kernel tuning configs for DeepSeek V3/R1 (sgl-project#5368) * Add H20 dtype fp8_w8a8 shared experts fused MoE kernel tuning configs for DeepSeek V3/R1 (sgl-project#5291) Co-authored-by: ximing.wxm <[email protected]> * [fix/misc] remove duplicate row in deepseek v2 model (sgl-project#5279) * chore: upgrade DeepGEMM (sgl-project#5395) * fix: update pr-test-sgl-kernel (sgl-project#5399) * kernel: support slightly faster merge_state_v2 cuda kernel (sgl-project#5381) * chore: bump sgl-kernel 0.0.9 (sgl-project#5400) * chore: upgrade sgl-kernel 0.0.9 (sgl-project#5401) * Tiny fix DeepseekScalingRotaryEmbedding always use forward_native (sgl-project#5406) * Fix bench_serving with random-ids (sgl-project#5214) * [misc] fix ci flaky case (sgl-project#5352) * [FIX] Fix concatenation error in capture_bs when open --disable-cuda-graph-padding and without MTP (sgl-project#5412) * Support dynamic connection and TP 16 (sgl-project#5351) Co-authored-by: luoyuan.luo <[email protected]> * Fix broadcast use cuda device lead to memory capacity unbalanced (sgl-project#5416) * [PD] Fix dynamic port support and MLA buffer for Mooncake (sgl-project#5415) Signed-off-by: Shangming Cai <[email protected]> Co-authored-by: ybyang <[email protected]> * Distinguish bootstrap key only in decode server (sgl-project#5422) * [PD] Remove unused bootstrap param and fix port table type (sgl-project#5423) * [minor] cleanup cmakelists.txt (sgl-project#5420) * bugfix: fix merge_state_v2 cuda graph (sgl-project#5419) * chore: bump sgl-kernel v0.0.9.post1 (sgl-project#5430) * fix: solve release issue (sgl-project#5434) * BLackwell cutlass mla: Add check for bad page size/block num combinations (sgl-project#5431) * feat: update model_specific_adjustment (sgl-project#5344) Co-authored-by: hebiao064 <[email protected]> * chore: upgrade sgl-kernel 0.0.9.post1 (sgl-project#5436) * Fix ignore_eos parameter when loading a chat template (sgl-project#5264) * add attention backend supporting matrix in the doc (sgl-project#5211) Co-authored-by: Stefan He <[email protected]> * Support BNB quantization for llama/mllama (sgl-project#5038) Co-authored-by: Yuhao Yang <[email protected]> * [Docs] Update start/install.md (sgl-project#5398) * [Minor] Move torch.compile patch to a better place (sgl-project#5397) * [Bug fix] need record start time in pd mode (sgl-project#5425) * Support MHA with chunked prefix cache for DeepSeek chunked prefill (sgl-project#5113) * chore: bump v0.4.5.post1 (sgl-project#5445) * Revert "[SW-226289] rebase sglang to tag v0.4.5 (sgl-project#12)" This reverts commit 0eac714. --------- Signed-off-by: Xinyuan Tong <[email protected]> Signed-off-by: Shangming Cai <[email protected]> Signed-off-by: Kay Yan <[email protected]> Signed-off-by: windsonsea <[email protected]> Co-authored-by: fzyzcjy <[email protected]> Co-authored-by: Lianmin Zheng <[email protected]> Co-authored-by: Juwan Yoo <[email protected]> Co-authored-by: Qingquan Song <[email protected]> Co-authored-by: Yineng Zhang <[email protected]> Co-authored-by: chaobo jia <[email protected]> Co-authored-by: rudy152 <[email protected]> Co-authored-by: Fr4nk1in <[email protected]> Co-authored-by: yinfan98 <[email protected]> Co-authored-by: Baizhou Zhang <[email protected]> Co-authored-by: Ke Bao <[email protected]> Co-authored-by: Yi Zhang <[email protected]> Co-authored-by: Adarsh Shirawalmath <[email protected]> Co-authored-by: Sleepcoo <[email protected]> Co-authored-by: SEPLOS <[email protected]> Co-authored-by: ch-wan <[email protected]> Co-authored-by: Zhiqiang Xie <[email protected]> Co-authored-by: JieXin Liang <[email protected]> Co-authored-by: Mick <[email protected]> Co-authored-by: Yuhong Guo <[email protected]> Co-authored-by: Jinyan Chen <[email protected]> Co-authored-by: laixinn <[email protected]> Co-authored-by: XinyuanTong <[email protected]> Co-authored-by: GeLee <[email protected]> Co-authored-by: Xiaoyu Zhang <[email protected]> Co-authored-by: hebiao064 <[email protected]> Co-authored-by: zcnrex <[email protected]> Co-authored-by: Kaiyu Yang <[email protected]> Co-authored-by: renxin <[email protected]> Co-authored-by: saltyfish66 <[email protected]> Co-authored-by: yuethe <[email protected]> Co-authored-by: simveit <[email protected]> Co-authored-by: Yifan Zhang <[email protected]> Co-authored-by: Ravi Theja <[email protected]> Co-authored-by: Ravi Theja Desetty <[email protected]> Co-authored-by: AniZpZ <[email protected]> Co-authored-by: 晟海 <[email protected]> Co-authored-by: Tommy Yang <[email protected]> Co-authored-by: Cheng Wan <[email protected]> Co-authored-by: inkcherry <[email protected]> Co-authored-by: mlmz <[email protected]> Co-authored-by: shuaills <[email protected]> Co-authored-by: Chang Su <[email protected]> Co-authored-by: fzyzcjy <[email protected]> Co-authored-by: HAI <[email protected]> Co-authored-by: tianhaoyu <[email protected]> Co-authored-by: DarkSharpness <[email protected]> Co-authored-by: Yun Dai <[email protected]> Co-authored-by: Hubert Lu <[email protected]> Co-authored-by: huangtingwei <[email protected]> Co-authored-by: kk <[email protected]> Co-authored-by: wunhuang <[email protected]> Co-authored-by: root <[email protected]> Co-authored-by: Yubo Wang <[email protected]> Co-authored-by: saienduri <[email protected]> Co-authored-by: DangKai <[email protected]> Co-authored-by: dangkai.dk <[email protected]> Co-authored-by: shangmingc <[email protected]> Co-authored-by: Ma Mingfei <[email protected]> Co-authored-by: Chunyuan WU <[email protected]> Co-authored-by: YanbingJiang <[email protected]> Co-authored-by: blzheng <[email protected]> Co-authored-by: Byron Hsu <[email protected]> Co-authored-by: Liangsheng Yin <[email protected]> Co-authored-by: zhaochenyang20 <[email protected]> Co-authored-by: Trevor Morris <[email protected]> Co-authored-by: Kay Yan <[email protected]> Co-authored-by: grimoire <[email protected]> Co-authored-by: HandH1998 <[email protected]> Co-authored-by: Zhaoyang Hao <[email protected]> Co-authored-by: Teng Ma <[email protected]> Co-authored-by: Shangming Cai <[email protected]> Co-authored-by: Xuchun Shang <[email protected]> Co-authored-by: Richard Zou <[email protected]> Co-authored-by: Elfie Guo <[email protected]> Co-authored-by: Michael Yao <[email protected]> Co-authored-by: Yusong Gao <[email protected]> Co-authored-by: Zhaoyi Li <[email protected]> Co-authored-by: lambert0312 <[email protected]> Co-authored-by: tianlian yi <[email protected]> Co-authored-by: Jin Pan <[email protected]> Co-authored-by: Jinn <[email protected]> Co-authored-by: yulei <[email protected]> Co-authored-by: Yongtong Wu <[email protected]> Co-authored-by: yhyang201 <[email protected]> Co-authored-by: ybyang <[email protected]> Co-authored-by: Ximingwang-09 <[email protected]> Co-authored-by: ximing.wxm <[email protected]> Co-authored-by: Yangcheng Li <[email protected]> Co-authored-by: DefTruth <[email protected]> Co-authored-by: Yuan Luo <[email protected]> Co-authored-by: luoyuan.luo <[email protected]> Co-authored-by: ybyang <[email protected]> Co-authored-by: mRSun15 <[email protected]> Co-authored-by: ryang <[email protected]> Co-authored-by: Yuhao Yang <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Motivation
Currently, Verlengine only supports engine-based rollout. However, server-based rollout not only supports all the features of engine-based rollout but also enables API call functionality, which is easier for continuous bataching. Therefore, we present a server-based rollout solution.
Modifications
My modifications focus on three main aspects:
Support for
update_weights_from_tensorin the HTTP server:To maintain compatibility with existing Verlengine features, I added support for the
update_weights_from_tensormethod in the HTTP server. This function receives metadata instead of the full tensor, allowing for fast and efficient weight updates.Implementation of
HttpServerEngineAdapter:I introduced the
HttpServerEngineAdapterclass, which mirrors all the methods used by the engine in theVerlengineclass. This allows it to function as a server-based rollout engine with full feature compatibility.Test file for validation:
A dedicated test file was added to ensure the correctness of the implementation and verify that all functionalities work as expected.
@jhinpan @zhaochenyang20
Checklist