From c654e5bbacbb2a80c42383758d97baca3624359f Mon Sep 17 00:00:00 2001 From: "junhao.zhang" Date: Fri, 31 Jul 2026 11:34:22 +0800 Subject: [PATCH 1/2] upload qwen3.5-27b and qwen3.6-27b models --- models/nlp/llm/qwen3.5-27b/vllm/README.md | 91 +++++++++++ models/nlp/llm/qwen3.5-27b/vllm/ci/prepare.sh | 17 +++ .../llm/qwen3.5-27b/vllm/offline_inference.py | 143 ++++++++++++++++++ models/nlp/llm/qwen3.6-27b/vllm/README.md | 97 ++++++++++++ models/nlp/llm/qwen3.6-27b/vllm/ci/prepare.sh | 17 +++ .../llm/qwen3.6-27b/vllm/offline_inference.py | 143 ++++++++++++++++++ 6 files changed, 508 insertions(+) create mode 100644 models/nlp/llm/qwen3.5-27b/vllm/README.md create mode 100644 models/nlp/llm/qwen3.5-27b/vllm/ci/prepare.sh create mode 100644 models/nlp/llm/qwen3.5-27b/vllm/offline_inference.py create mode 100644 models/nlp/llm/qwen3.6-27b/vllm/README.md create mode 100644 models/nlp/llm/qwen3.6-27b/vllm/ci/prepare.sh create mode 100644 models/nlp/llm/qwen3.6-27b/vllm/offline_inference.py diff --git a/models/nlp/llm/qwen3.5-27b/vllm/README.md b/models/nlp/llm/qwen3.5-27b/vllm/README.md new file mode 100644 index 00000000..dd243cea --- /dev/null +++ b/models/nlp/llm/qwen3.5-27b/vllm/README.md @@ -0,0 +1,91 @@ +# Qwen3.5-27B (vLLM) + +## Model Description + +Qwen3.5-27B is a multimodal dialogue model of the Qwen3.5 series (architecture `Qwen3_5ForConditionalGeneration`). It supports text / image / video input and switchable thinking mode. BF16 native weights, about 52 GB / 11 safetensors shards (`00001-of-00011` … `00011-of-00011`), hidden size 5120, 64 layers, native context 262144. + +## Supported Environments + +| GPU | [IXUCA SDK](https://gitee.com/deep-spark/deepspark#%E5%A4%A9%E6%95%B0%E6%99%BA%E7%AE%97%E8%BD%AF%E4%BB%B6%E6%A0%88-ixuca) | Release | Branch | +| :----: | :----: | :----: | :----: | +| BI-V150 | dev-only | 26.09 | — | + +> **Note:** 请切换到与您的 SDK 版本对应的 Release 分支进行测试。请勿直接在 master 分支上运行测试，因为 master 分支可能包含与您的本地 SDK 版本不兼容的最新更改。 + +## Model Preparation + +### Prepare Resources + +- Model: + +```bash +cd models/nlp/llm/qwen3.5-27b/vllm +mkdir -p data/qwen3 +ln -s /path/to/Qwen3.5-27B ./data/qwen3 +``` + +### Install Dependencies + +In order to run the model smoothly, you need to get the sdk from [resource center](https://support.iluvatar.com/#/ProductLine?id=2) of Iluvatar CoreX official website. + +## Model Inference + +### Inference with offline + +```bash +python3 offline_inference.py \ + --model ./data/qwen3/Qwen3.5-27B \ + --max-tokens 256 -tp 4 \ + --trust-remote-code --temperature 0.0 \ + --max-model-len 4096 +``` + +### Inference with serve + +```bash +python3 -m vllm.entrypoints.openai.api_server \ + --model ./data/qwen3/Qwen3.5-27B \ + --served-model-name Qwen3.5-27B \ + --trust-remote-code \ + --tensor-parallel-size 4 \ + --max-model-len 40960 \ + --gpu-memory-utilization 0.9 \ + --reasoning-parser qwen3 \ + --port 8000 +``` + +> BF16 dense model. Do **not** set `VLLM_W8A8_MOE_USE_W4A8`. + +## Model Results + +### Benchmarking vLLM + +```bash +export CUDA_VISIBLE_DEVICES=0,1,3,4 +# If sonnet.txt is missing: +# curl -fsSL -o /tmp/sonnet.txt \ +# https://raw.githubusercontent.com/vllm-project/vllm/main/benchmarks/sonnet.txt + +vllm bench throughput \ + --model ./data/qwen3/Qwen3.5-27B \ + --dataset-name sonnet \ + --dataset-path /tmp/sonnet.txt \ + --num-prompts 10 \ + --trust-remote-code \ + --tensor-parallel-size 4 \ + --max-model-len 4096 \ + --gpu-memory-utilization 0.9 +``` + +### Benchmarking Results + +| Model | Precision | QPS | Total TPS | Output TPS | +| :----: | :----: | :----: | :----: | :----: | +| Qwen3.5-27B | BF16 | — | — | — | + +## References + +- [Qwen3.5](https://github.com/QwenLM/Qwen3.5) +- [vLLM](https://github.com/vllm-project/vllm) +- [DeepSparkInference · Qwen3-8B](https://gitee.com/deep-spark/deepsparkinference/tree/master/models/nlp/llm/qwen3-8b/vllm) +- [DeepSparkInference · Qwen2-7B](https://gitee.com/deep-spark/deepsparkinference/tree/master/models/nlp/llm/qwen2-7b/vllm) diff --git a/models/nlp/llm/qwen3.5-27b/vllm/ci/prepare.sh b/models/nlp/llm/qwen3.5-27b/vllm/ci/prepare.sh new file mode 100644 index 00000000..4a733ba9 --- /dev/null +++ b/models/nlp/llm/qwen3.5-27b/vllm/ci/prepare.sh @@ -0,0 +1,17 @@ +#!/bin/bash +# Copyright (c) 2025, Shanghai Iluvatar CoreX Semiconductor Co., Ltd. +# All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); you may +# not use this file except in compliance with the License. You may obtain +# a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +set -x diff --git a/models/nlp/llm/qwen3.5-27b/vllm/offline_inference.py b/models/nlp/llm/qwen3.5-27b/vllm/offline_inference.py new file mode 100644 index 00000000..ca7b4d78 --- /dev/null +++ b/models/nlp/llm/qwen3.5-27b/vllm/offline_inference.py @@ -0,0 +1,143 @@ +# Copyright (c) 2025, Shanghai Iluvatar CoreX Semiconductor Co., Ltd. +# All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); you may +# not use this file except in compliance with the License. You may obtain +# a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from vllm import LLM, EngineArgs, SamplingParams +import sys +from pathlib import Path +import os +import argparse as _argparse +import dataclasses + +# ====== PATCH: 兼容旧版 argparse 不支持 'deprecated' ====== +_original_add_argument = _argparse._ArgumentGroup.add_argument + +def _patched_add_argument(self, *args, **kwargs): + kwargs.pop('deprecated', None) + return _original_add_argument(self, *args, **kwargs) + +_argparse._ArgumentGroup.add_argument = _patched_add_argument +# ========================================================= +sys.path.append(str(Path(__file__).resolve().parent.parent.parent)) +import argparse +import dataclasses +import inspect +import logging +import time + +import torch +from utils import load_chat_template, sampling_add_cli_args +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument("--chat_template", type=str, default=None) + parser.add_argument( + "--remove_chat_template", + default=False, + action="store_true", + help="pass this if you are not use a chat model", + ) + parser = EngineArgs.add_cli_args(parser) + parser = sampling_add_cli_args(parser) + args = parser.parse_args() + + engine_args = EngineArgs.from_cli_args(args) + engine_params = dataclasses.asdict(engine_args) + sampling_args = [ + param.name + for param in list( + inspect.signature(SamplingParams).parameters.values() + ) + ] + sampling_params = { + attr: getattr(args, attr) for attr in sampling_args if args.__contains__(attr) + } + + model_name = os.path.dirname(args.model).rsplit("/")[-1] + + # Sample prompts. + prompts = ["哪些迹象可能表明一个人正在经历焦虑?", "描述一下如何制作芝士披萨。", "写一篇有关5G网络研发的综述文章。"] + + # Create a sampling params object. + sampling_params = SamplingParams(**sampling_params) + + # Create an LLM. + llm = LLM(**engine_params) + + # process chat template + if args.remove_chat_template: + if "chat" in model_name.lower(): + logging.warning( + f"The model name from model path is {model_name}, so we guess you are using the chat model and the additional processing is required for the input prompt. " + f"If the result is not quite correct, please ensure you do not pass --remove_chat_template in CLI." + ) + prompts_new = prompts + else: + # Build chat model promopt + logging.warning( + "If you are using a non chat model, please pass the --remove_chat_template in CLI." + ) + # Try use transformers's apply_chat_template, if chat_template is None, will use defalut template. + # For some old models, the default template may cause bad answers. we don't consider this situation, + # because the Transformers team is advancing the chat template. For more informatino about it, + # please refer to https://huggingface.co/docs/transformers/main/chat_templating + try: + load_chat_template(llm.get_tokenizer(), args.chat_template) + prompts_new = [] + for prompt in prompts: + messages = [{"role": "user", "content": prompt}] + text = llm.get_tokenizer().apply_chat_template( + messages, tokenize=False, add_generation_prompt=True + ) + prompts_new.append(text) + except: + logging.warning( + "use tokenizer apply_chat_template function failed, may because of low transformers version...(try use transformers>=4.34.0)" + ) + + prompts_new = prompts + + # Generate texts from the prompts. The output is a list of RequestOutput objects + # that contain the prompt, generated text, and other information. + outputs = ( + llm.generate(prompts_new, sampling_params, use_tqdm=False) + if isinstance(prompts_new[0], str) + else llm.generate( + sampling_params=sampling_params, + prompt_token_ids=prompts_new, + use_tqdm=False, + ) + ) + torch.cuda.synchronize() + + start_time = time.perf_counter() + outputs = ( + llm.generate(prompts_new, sampling_params) + if isinstance(prompts_new[0], str) + else llm.generate(sampling_params=sampling_params, prompt_token_ids=prompts_new) + ) + torch.cuda.synchronize() + end_time = time.perf_counter() + duration_time = end_time - start_time + + num_tokens = 0 + # Print the outputs. + for i, output in enumerate(outputs): + prompt = prompts[i] # show the origin prompt. actully prompt is "output.prompt" + generated_text = output.outputs[0].text + + num_tokens += len(output.outputs[0].token_ids) + print(f"Prompt: {prompt}\nGenerated text: {generated_text} \n") + num_requests = len(prompts) # 请求的数量 + qps = num_requests / duration_time + print(f"requests: {num_requests}, QPS: {qps}, tokens: {num_tokens}, Token/s: {num_tokens/duration_time}") diff --git a/models/nlp/llm/qwen3.6-27b/vllm/README.md b/models/nlp/llm/qwen3.6-27b/vllm/README.md new file mode 100644 index 00000000..b76aa835 --- /dev/null +++ b/models/nlp/llm/qwen3.6-27b/vllm/README.md @@ -0,0 +1,97 @@ +# Qwen3.6-27B (vLLM) + +## Model Description + +Qwen3.6-27B is a multimodal dialogue model of the Qwen3.6 series (architecture `Qwen3_5ForConditionalGeneration`, `model_type=qwen3_5`). It supports text / image / video input and switchable thinking mode. BF16 native weights, about 52 GB (`model.safetensors.index.json` reports 51.75 GiB / 1199 tensors) / 15 safetensors shards (`model-00001-of-00015` … `model-00015-of-00015`). Hidden size 5120, 64 layers, 24 attention heads / 4 KV heads, FFN intermediate size 17408, vocab size 248320, native context 262144, with vision tower. + +## Supported Environments + +| GPU | [IXUCA SDK](https://gitee.com/deep-spark/deepspark#%E5%A4%A9%E6%95%B0%E6%99%BA%E7%AE%97%E8%BD%AF%E4%BB%B6%E6%A0%88-ixuca) | Release | Branch | +| :----: | :----: | :----: | :----: | +| BI-V150 | dev-only | 26.09 | — | + +> **Note:** 请切换到与您的 SDK 版本对应的 Release 分支进行测试。请勿直接在 master 分支上运行测试，因为 master 分支可能包含与您的本地 SDK 版本不兼容的最新更改。 + +## Model Preparation + +### Prepare Resources + +- Model: + +```bash +cd models/nlp/llm/qwen3.6-27b/vllm +mkdir -p data/qwen3 +ln -s /path/to/Qwen3.6-27B ./data/qwen3 +``` + +### Install Dependencies + +In order to run the model smoothly, you need to get the sdk from [resource center](https://support.iluvatar.com/#/ProductLine?id=2) of Iluvatar CoreX official website. + +## Model Inference + +### Inference with offline + +```bash +python3 offline_inference.py \ + --model ./data/qwen3/Qwen3.6-27B \ + --max-tokens 256 -tp 4 \ + --trust-remote-code --temperature 0.0 \ + --max-model-len 4096 +``` + +### Inference with serve + +```bash +python3 -m vllm.entrypoints.openai.api_server \ + --model ./data/qwen3/Qwen3.6-27B \ + --served-model-name Qwen3.6-27B \ + --trust-remote-code \ + --tensor-parallel-size 4 \ + --max-model-len 40960 \ + --gpu-memory-utilization 0.9 \ + --reasoning-parser qwen3 \ + --port 8000 +``` + +> BF16 dense model. Do **not** set `VLLM_W8A8_MOE_USE_W4A8` (that flag is only for MoE W4A8 checkpoints such as Qwen3.6-35B-A3B-W4A8). + +## Model Results + +### Benchmarking vLLM + +```bash +# If sonnet.txt is missing: +# curl -fsSL -o /tmp/sonnet.txt \ +# https://raw.githubusercontent.com/vllm-project/vllm/main/benchmarks/sonnet.txt + +vllm bench throughput \ + --model ./data/qwen3/Qwen3.6-27B \ + --dataset-name sonnet \ + --dataset-path /tmp/sonnet.txt \ + --num-prompts 10 \ + --trust-remote-code \ + --tensor-parallel-size 4 \ + --max-model-len 4096 \ + --gpu-memory-utilization 0.9 +``` + +### Benchmarking Results + +| Model | Precision | QPS | Total TPS | Output TPS | +| :----: | :----: | :----: | :----: | :----: | +| Qwen3.6-27B | BF16 | — | — | — | + +Horizontal comparison under the same environment and load: + +| Model | Precision | QPS | Total TPS | Output TPS | +| :----: | :----: | :----: | :----: | :----: | +| Qwen3.6-27B | BF16 | — | — | — | +| Qwen3.6-27B-W8A8 | INT8 (W8A8) | — | — | — | +| Qwen3.5-27B | BF16 | — | — | — | + +## References + +- [Qwen3.5](https://github.com/QwenLM/Qwen3.5) +- [vLLM](https://github.com/vllm-project/vllm) +- [DeepSparkInference · Qwen3-8B](https://gitee.com/deep-spark/deepsparkinference/tree/master/models/nlp/llm/qwen3-8b/vllm) diff --git a/models/nlp/llm/qwen3.6-27b/vllm/ci/prepare.sh b/models/nlp/llm/qwen3.6-27b/vllm/ci/prepare.sh new file mode 100644 index 00000000..4a733ba9 --- /dev/null +++ b/models/nlp/llm/qwen3.6-27b/vllm/ci/prepare.sh @@ -0,0 +1,17 @@ +#!/bin/bash +# Copyright (c) 2025, Shanghai Iluvatar CoreX Semiconductor Co., Ltd. +# All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); you may +# not use this file except in compliance with the License. You may obtain +# a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +set -x diff --git a/models/nlp/llm/qwen3.6-27b/vllm/offline_inference.py b/models/nlp/llm/qwen3.6-27b/vllm/offline_inference.py new file mode 100644 index 00000000..ca7b4d78 --- /dev/null +++ b/models/nlp/llm/qwen3.6-27b/vllm/offline_inference.py @@ -0,0 +1,143 @@ +# Copyright (c) 2025, Shanghai Iluvatar CoreX Semiconductor Co., Ltd. +# All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); you may +# not use this file except in compliance with the License. You may obtain +# a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from vllm import LLM, EngineArgs, SamplingParams +import sys +from pathlib import Path +import os +import argparse as _argparse +import dataclasses + +# ====== PATCH: 兼容旧版 argparse 不支持 'deprecated' ====== +_original_add_argument = _argparse._ArgumentGroup.add_argument + +def _patched_add_argument(self, *args, **kwargs): + kwargs.pop('deprecated', None) + return _original_add_argument(self, *args, **kwargs) + +_argparse._ArgumentGroup.add_argument = _patched_add_argument +# ========================================================= +sys.path.append(str(Path(__file__).resolve().parent.parent.parent)) +import argparse +import dataclasses +import inspect +import logging +import time + +import torch +from utils import load_chat_template, sampling_add_cli_args +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument("--chat_template", type=str, default=None) + parser.add_argument( + "--remove_chat_template", + default=False, + action="store_true", + help="pass this if you are not use a chat model", + ) + parser = EngineArgs.add_cli_args(parser) + parser = sampling_add_cli_args(parser) + args = parser.parse_args() + + engine_args = EngineArgs.from_cli_args(args) + engine_params = dataclasses.asdict(engine_args) + sampling_args = [ + param.name + for param in list( + inspect.signature(SamplingParams).parameters.values() + ) + ] + sampling_params = { + attr: getattr(args, attr) for attr in sampling_args if args.__contains__(attr) + } + + model_name = os.path.dirname(args.model).rsplit("/")[-1] + + # Sample prompts. + prompts = ["哪些迹象可能表明一个人正在经历焦虑?", "描述一下如何制作芝士披萨。", "写一篇有关5G网络研发的综述文章。"] + + # Create a sampling params object. + sampling_params = SamplingParams(**sampling_params) + + # Create an LLM. + llm = LLM(**engine_params) + + # process chat template + if args.remove_chat_template: + if "chat" in model_name.lower(): + logging.warning( + f"The model name from model path is {model_name}, so we guess you are using the chat model and the additional processing is required for the input prompt. " + f"If the result is not quite correct, please ensure you do not pass --remove_chat_template in CLI." + ) + prompts_new = prompts + else: + # Build chat model promopt + logging.warning( + "If you are using a non chat model, please pass the --remove_chat_template in CLI." + ) + # Try use transformers's apply_chat_template, if chat_template is None, will use defalut template. + # For some old models, the default template may cause bad answers. we don't consider this situation, + # because the Transformers team is advancing the chat template. For more informatino about it, + # please refer to https://huggingface.co/docs/transformers/main/chat_templating + try: + load_chat_template(llm.get_tokenizer(), args.chat_template) + prompts_new = [] + for prompt in prompts: + messages = [{"role": "user", "content": prompt}] + text = llm.get_tokenizer().apply_chat_template( + messages, tokenize=False, add_generation_prompt=True + ) + prompts_new.append(text) + except: + logging.warning( + "use tokenizer apply_chat_template function failed, may because of low transformers version...(try use transformers>=4.34.0)" + ) + + prompts_new = prompts + + # Generate texts from the prompts. The output is a list of RequestOutput objects + # that contain the prompt, generated text, and other information. + outputs = ( + llm.generate(prompts_new, sampling_params, use_tqdm=False) + if isinstance(prompts_new[0], str) + else llm.generate( + sampling_params=sampling_params, + prompt_token_ids=prompts_new, + use_tqdm=False, + ) + ) + torch.cuda.synchronize() + + start_time = time.perf_counter() + outputs = ( + llm.generate(prompts_new, sampling_params) + if isinstance(prompts_new[0], str) + else llm.generate(sampling_params=sampling_params, prompt_token_ids=prompts_new) + ) + torch.cuda.synchronize() + end_time = time.perf_counter() + duration_time = end_time - start_time + + num_tokens = 0 + # Print the outputs. + for i, output in enumerate(outputs): + prompt = prompts[i] # show the origin prompt. actully prompt is "output.prompt" + generated_text = output.outputs[0].text + + num_tokens += len(output.outputs[0].token_ids) + print(f"Prompt: {prompt}\nGenerated text: {generated_text} \n") + num_requests = len(prompts) # 请求的数量 + qps = num_requests / duration_time + print(f"requests: {num_requests}, QPS: {qps}, tokens: {num_tokens}, Token/s: {num_tokens/duration_time}") -- Gitee From e321fc6102f9f3d86af5ce328655cfad44f62809 Mon Sep 17 00:00:00 2001 From: "junhao.zhang" Date: Wed, 5 Aug 2026 13:43:01 +0800 Subject: [PATCH 2/2] Fixed some issues with the description --- models/nlp/llm/qwen3.5-27b/vllm/README.md | 4 ++-- models/nlp/llm/qwen3.5-27b/vllm/ci/prepare.sh | 2 +- models/nlp/llm/qwen3.6-27b/vllm/README.md | 4 ++-- models/nlp/llm/qwen3.6-27b/vllm/ci/prepare.sh | 2 +- 4 files changed, 6 insertions(+), 6 deletions(-) diff --git a/models/nlp/llm/qwen3.5-27b/vllm/README.md b/models/nlp/llm/qwen3.5-27b/vllm/README.md index dd243cea..ab9e0a72 100644 --- a/models/nlp/llm/qwen3.5-27b/vllm/README.md +++ b/models/nlp/llm/qwen3.5-27b/vllm/README.md @@ -8,9 +8,9 @@ Qwen3.5-27B is a multimodal dialogue model of the Qwen3.5 series (architecture ` | GPU | [IXUCA SDK](https://gitee.com/deep-spark/deepspark#%E5%A4%A9%E6%95%B0%E6%99%BA%E7%AE%97%E8%BD%AF%E4%BB%B6%E6%A0%88-ixuca) | Release | Branch | | :----: | :----: | :----: | :----: | -| BI-V150 | dev-only | 26.09 | — | +| MR-V100 | dev-only | 26.09 | — | -> **Note:** 请切换到与您的 SDK 版本对应的 Release 分支进行测试。请勿直接在 master 分支上运行测试，因为 master 分支可能包含与您的本地 SDK 版本不兼容的最新更改。 +> **Note:** 请切换到 release/26.09 分支进行测试。请勿直接在 master 分支上运行测试，因为 master 分支可能包含与您的本地 SDK 版本不兼容的最新更改。 ## Model Preparation diff --git a/models/nlp/llm/qwen3.5-27b/vllm/ci/prepare.sh b/models/nlp/llm/qwen3.5-27b/vllm/ci/prepare.sh index 4a733ba9..0231c83d 100644 --- a/models/nlp/llm/qwen3.5-27b/vllm/ci/prepare.sh +++ b/models/nlp/llm/qwen3.5-27b/vllm/ci/prepare.sh @@ -1,5 +1,5 @@ #!/bin/bash -# Copyright (c) 2025, Shanghai Iluvatar CoreX Semiconductor Co., Ltd. +# Copyright (c) 2026, Shanghai Iluvatar CoreX Semiconductor Co., Ltd. # All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); you may diff --git a/models/nlp/llm/qwen3.6-27b/vllm/README.md b/models/nlp/llm/qwen3.6-27b/vllm/README.md index b76aa835..fb8b8d24 100644 --- a/models/nlp/llm/qwen3.6-27b/vllm/README.md +++ b/models/nlp/llm/qwen3.6-27b/vllm/README.md @@ -8,9 +8,9 @@ Qwen3.6-27B is a multimodal dialogue model of the Qwen3.6 series (architecture ` | GPU | [IXUCA SDK](https://gitee.com/deep-spark/deepspark#%E5%A4%A9%E6%95%B0%E6%99%BA%E7%AE%97%E8%BD%AF%E4%BB%B6%E6%A0%88-ixuca) | Release | Branch | | :----: | :----: | :----: | :----: | -| BI-V150 | dev-only | 26.09 | — | +| MR-V100 | dev-only | 26.09 | — | -> **Note:** 请切换到与您的 SDK 版本对应的 Release 分支进行测试。请勿直接在 master 分支上运行测试，因为 master 分支可能包含与您的本地 SDK 版本不兼容的最新更改。 +> **Note:** 请切换到 release/26.09 分支进行测试。请勿直接在 master 分支上运行测试，因为 master 分支可能包含与您的本地 SDK 版本不兼容的最新更改。 ## Model Preparation diff --git a/models/nlp/llm/qwen3.6-27b/vllm/ci/prepare.sh b/models/nlp/llm/qwen3.6-27b/vllm/ci/prepare.sh index 4a733ba9..0231c83d 100644 --- a/models/nlp/llm/qwen3.6-27b/vllm/ci/prepare.sh +++ b/models/nlp/llm/qwen3.6-27b/vllm/ci/prepare.sh @@ -1,5 +1,5 @@ #!/bin/bash -# Copyright (c) 2025, Shanghai Iluvatar CoreX Semiconductor Co., Ltd. +# Copyright (c) 2026, Shanghai Iluvatar CoreX Semiconductor Co., Ltd. # All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); you may -- Gitee