Skip to content

[DOC] enable both flashcomm1 and cudagraph #6807

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
whx-sjtu merged 1 commit into from
Feb 27, 2026
Merged

[DOC] enable both flashcomm1 and cudagraph #6807

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 15 additions & 7 deletions docs/source/tutorials/models/DeepSeek-V3.2.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -140,6 +140,7 @@ export VLLM_USE_V1=1
export HCCL_BUFFSIZE=200
export VLLM_ASCEND_ENABLE_MLAPO=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export VLLM_ASCEND_ENABLE_FLASHCOMM1=1

vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V3.2-W8A8 \
--host 0.0.0.0 \
Expand All @@ -157,6 +158,7 @@ vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V3.2-W8A8 \
--no-enable-prefix-caching \
--gpu-memory-utilization 0.92 \
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
--additional-config '{"layer_sharding": ["q_b_proj", "o_proj"]}' \
--speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp"}'

```
Expand Down Expand Up @@ -197,6 +199,7 @@ export VLLM_USE_V1=1
export HCCL_BUFFSIZE=200
export VLLM_ASCEND_ENABLE_MLAPO=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export VLLM_ASCEND_ENABLE_FLASHCOMM1=1

vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V3.2-W8A8 \
--host 0.0.0.0 \
Expand All @@ -217,6 +220,7 @@ vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V3.2-W8A8 \
--no-enable-prefix-caching \
--gpu-memory-utilization 0.92 \
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
--additional-config '{"layer_sharding": ["q_b_proj", "o_proj"]}' \
--speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp"}'
```

Expand Down Expand Up @@ -244,6 +248,7 @@ export VLLM_USE_V1=1
export HCCL_BUFFSIZE=200
export VLLM_ASCEND_ENABLE_MLAPO=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export VLLM_ASCEND_ENABLE_FLASHCOMM1=1

vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V3.2-W8A8 \
--host 0.0.0.0 \
Expand All @@ -266,6 +271,7 @@ vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V3.2-W8A8 \
--no-enable-prefix-caching \
--gpu-memory-utilization 0.92 \
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
--additional-config '{"layer_sharding": ["q_b_proj", "o_proj"]}' \
--speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp"}'
```

Expand Down Expand Up @@ -301,7 +307,7 @@ export VLLM_ASCEND_ENABLE_FLASHCOMM1=1
export HCCL_CONNECT_TIMEOUT=120
export HCCL_INTRA_PCIE_ENABLE=1
export HCCL_INTRA_ROCE_ENABLE=0

export VLLM_ASCEND_ENABLE_FLASHCOMM1=1
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist Bot Feb 25, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

The environment variable VLLM_ASCEND_ENABLE_FLASHCOMM1 is already exported on line 306. This duplicate export is redundant and should be removed for clarity.


vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V3.2-W8A8 \
--host 0.0.0.0 \
Expand All @@ -321,8 +327,9 @@ vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V3.2-W8A8 \
--trust-remote-code \
--no-enable-prefix-caching \
--gpu-memory-utilization 0.92 \
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY", "cudagraph_capture_sizes":[3,6,9,12,15,18,21,24,27,30,33,36,39,42,45,48]}' \
--speculative-config '{"num_speculative_tokens": 2, "method": "deepseek_mtp"}'
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY", "cudagraph_capture_sizes":[8, 16, 24, 32, 40, 48]}' \
--additional-config '{"layer_sharding": ["q_b_proj", "o_proj"]}' \
--speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp"}'

```

Expand Down Expand Up @@ -354,7 +361,7 @@ export VLLM_ASCEND_ENABLE_FLASHCOMM1=1
export HCCL_CONNECT_TIMEOUT=120
export HCCL_INTRA_PCIE_ENABLE=1
export HCCL_INTRA_ROCE_ENABLE=0

export VLLM_ASCEND_ENABLE_FLASHCOMM1=1
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist Bot Feb 25, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

The environment variable VLLM_ASCEND_ENABLE_FLASHCOMM1 is already exported on line 360. This duplicate export is redundant and should be removed for clarity.


vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V3.2-W8A8 \
--host 0.0.0.0 \
Expand All @@ -376,8 +383,9 @@ vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V3.2-W8A8 \
--trust-remote-code \
--no-enable-prefix-caching \
--gpu-memory-utilization 0.92 \
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY", "cudagraph_capture_sizes":[3,6,9,12,15,18,21,24,27,30,33,36,39,42,45,48]}' \
--speculative-config '{"num_speculative_tokens": 2, "method": "deepseek_mtp"}'
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY", "cudagraph_capture_sizes":[8, 16, 24, 32, 40, 48]}' \
--additional-config '{"layer_sharding": ["q_b_proj", "o_proj"]}' \
--speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp"}'

```

Expand Down Expand Up @@ -832,7 +840,7 @@ python launch_online_dp.py --dp-size 8 --tp-size 4 --dp-size-local 4 --dp-rank-s

### Request Forwarding

To set up request forwarding, run the following script on any machine :download:`load_balance_proxy_server_example.py <examples/disaggregated_prefill_v1/load_balance_proxy_server_example.py>`
To set up request forwarding, run the following script on any machine. You can get the proxy program in the repository's examples: [load_balance_proxy_server_example.py](https://github.com/vllm-project/vllm-ascend/blob/main/examples/disaggregated_prefill_v1/load_balance_proxy_server_example.py)

```shell
unset http_proxy
Expand Down
Loading