[Doc] Add layer_sharding additional config for DeepSeek-V3.2-W8A8 (#5921)
### What this PR does / why we need it?
#### Documentation Improvements
New Configuration: Added the layer_sharding parameter to the
DeepSeek-V3.2-W8A8 deployment tutorial. This guides users to include
`["q_b_proj", "o_proj"]` in their prefill node setup for better resource
utilization.
#### CI and Testing Updates
Test Config Update: Updated the multi-node E2E test configuration file:
tests/e2e/nightly/multi_node/config/DeepSeek-V3_2-W8A8-A3-dual-nodes.yaml.
including disable `FLASHCOMM` and enable `FULL_DECODE_ONLY` and update
performance baseline.
### Does this PR introduce any user-facing change?
Yes. The documentation now recommends a more optimized startup command
for DeepSeek-V3.2-W8A8. Users following the updated tutorial will see
improved performance in multi-node PD disaggregation environments.
### How was this patch tested?
CI Validation: The updated E2E test configuration has been verified
through the nightly CI pipeline.
Environment: * vLLM version: v0.13.0
Base Commit:
[11b6af5](11b6af5280)
Hardware: Ascend A3/A2 multi-node cluster.
---------
Signed-off-by: guozr <guozr1997@hotmail.com>
Co-authored-by: guozr <guozr1997@hotmail.com>
This commit is contained in:
@@ -313,6 +313,7 @@ Before you start, please
|
||||
--quantization ascend \
|
||||
--enforce-eager \
|
||||
--no-enable-prefix-caching \
|
||||
--additional-config '{"layer_sharding": ["q_b_proj", "o_proj"]}' \
|
||||
--kv-transfer-config \
|
||||
'{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
@@ -388,6 +389,7 @@ Before you start, please
|
||||
--quantization ascend \
|
||||
--enforce-eager \
|
||||
--no-enable-prefix-caching \
|
||||
--additional-config '{"layer_sharding": ["q_b_proj", "o_proj"]}' \
|
||||
--kv-transfer-config \
|
||||
'{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
|
||||
Reference in New Issue
Block a user