-
Notifications
You must be signed in to change notification settings - Fork 257
[perf]Support MOE Multi-stream in Deepseek #947
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.
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
Conversation
global_bs = 0 | ||
moe_expert_num = len(expert_map) | ||
# hidden_states = hidden_states.bfloat16() | ||
kwargs = { | ||
kwargs1 = { |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
rename to a readable name
vllm_ascend/envs.py
Outdated
@@ -36,6 +36,8 @@ | |||
lambda: bool(int(os.getenv("COMPILE_CUSTOM_KERNELS", "1"))), | |||
"VLLM_ENABLE_MC2": | |||
lambda: bool(int(os.getenv("VLLM_ENABLE_MC2", '0'))), | |||
"VLLM_ENABLE_CV_PARALLEL": |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
use additional_config instead of env, since this change is only used for torchair GE mode. like #839 does, there are another 3 new config option coming.
how about
{
"additional_config": {
"torchair_graph_config": {
"enable": True,
"enable_cv_parallet": True,
"batch_sizes": "12345",
"batch_sizes_init": True
}
}
}
cc @zzzzwwjj
And don't forget add e2e test. The model weight is here: https://www.modelscope.cn/models/vllm-ascend/DeepSeek-V2-Lite-W8A8 take https://github.com/vllm-project/vllm-ascend/blob/main/tests/multicard/test_offline_inference_distributed.py#L49 as an example |
Signed-off-by: David9857 <985700846@qq.com> use additional_config to enable cv parallel Signed-off-by: David9857 <985700846@qq.com> rename kwargs1 in fused_experts_with_mc2 Signed-off-by: David9857 <985700846@qq.com>
Signed-off-by: David9857 <985700846@qq.com>
This pull request has conflicts, please resolve those before we can evaluate the pull request. |
This pull request has conflicts, please resolve those before we can evaluate the pull request. |
vllm_ascend/models/deepseek_v2.py
Outdated
@@ -179,6 +180,12 @@ def __init__( | |||
else: | |||
self.gate.e_score_correction_bias = None | |||
|
|||
self.enable_cv_parallel = False | |||
additional_config = get_current_vllm_config().additional_config |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
please use ascend_config instead now. Note that doc should be updated at the same time.
Signed-off-by: David9857 <985700846@qq.com>
Signed-off-by: David9857 <985700846@qq.com>
This pull request has conflicts, please resolve those before we can evaluate the pull request. |
Signed-off-by: David9857 <985700846@qq.com> bugfix Signed-off-by: David9857 <985700846@qq.com>
Signed-off-by: David9857 <985700846@qq.com>
What this PR does / why we need it?
Support MOE inner Multi-stream for Deepseek.
This feature requires graph mode with mc2 enabled.
Does this PR introduce any user-facing change?
How was this patch tested?