A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Doc] Update Batch-level DP docs (#25757)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Sep 26, 2025 at 17:37 UTC
633f943e30a4444d890d26b81850f7217736f840
1 file changed
+6
-5
docs/configuration/optimization.md
+6
-5
@@ -139,9 +139,9 @@ there is relatively little gain from TP. On the other hand, TP incurs significan
139
overhead because of all-reduce being performed after every layer.
140
141
Given this, it may be advantageous to instead shard the batched input data using TP, essentially
142
-performing batch-level DP. This has been shown to improve the throughput by around 10% for
142
+performing batch-level DP. This has been shown to improve the throughput and TTFT by around 10% for
143
`tensor_parallel_size=8`. For vision encoders that use hardware-unoptimized Conv3D operations,
144
-batch-level DP can provide another 40% increase to throughput compared to regular TP.
144
+batch-level DP can provide another 40% improvement compared to regular TP.
145
146
Nevertheless, since the weights of the multi-modal encoder are replicated across each TP rank,
147
there will be a minor increase in memory consumption and may cause OOM if you can barely fit the model already.
@@ -172,14 +172,15 @@ Batch-level DP needs to be implemented on a per-model basis,
172
and enabled by setting `supports_encoder_tp_data = True` in the model class.
173
Regardless, you need to set `mm_encoder_tp_mode="data"` in engine arguments to use this feature.
174
175
-Known supported models:
175
+Known supported models (with corresponding benchmarks):
176
177
-- GLM-4.5V GLM-4.1V (<gh-pr:23168>)
177
+- dots_ocr (<gh-pr:25466>)
178
+- GLM-4.1V or above (<gh-pr:23168>)
179
- InternVL (<gh-pr:23909>)
180
- Kimi-VL (<gh-pr:23817>)
181
- Llama4 (<gh-pr:18368>)
182
- MiniCPM-V-2.5 or above (<gh-pr:23327>, <gh-pr:23948>)
182
-- Qwen2.5-VL (<gh-pr:22742>)
183
+- Qwen2-VL or above (<gh-pr:22742>, <gh-pr:24955>, <gh-pr:25445>)
184
- Step3 (<gh-pr:22697>)
185
186
## Input Processing