anodyine/vllm-omni
1
0 Fork 0
Code Issues 4 Pull Requests 2 Actions Packages Projects Releases Wiki Activity

[HunyuanImage3.0][Performance][Optimazation]Adjust perf config (#4149)

Browse Source 
Signed-off-by: dengyunyang <584797741@qq.com>
This commit is contained in:
dengyunyang authored and GitHub committed 2026-06-04 22:54:46 +08:00
1 parent 519cf3ecf1
commit 469c1dc87e
2 files changed
+15 -11
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

tests/dfx/perf/tests/test_hunyuan_image3_it2i.json
+11 -10
View File
@@ -1,7 +1,7 @@
[
{
"test_name": "test_hunyuan_image3_ti2i_tp4_tp4_inline",
"description": "AR (TP=4) + DiT (TP=4) - using inline dataset and deploy-config",
"test_name": "test_hunyuan_image3_ti2i_inline",
"description": "AR (TP=2) + 2 * DiT (TP=2) - using inline dataset and deploy-config",
"server_type": "vllm-omni",
"benchmark_endpoint": "/v1/images/edits",
"server_params": {
@@ -21,7 +21,7 @@
"zmq_port": 50051,
"protocol": "rdma",
"device_name": "",
"memory_pool_size": 4294967296,
"memory_pool_size": 17179869184,
"memory_pool_device": "cpu"
}
}
@@ -32,13 +32,13 @@
"is_comprehension": true,
"final_output": true,
"final_output_type": "text",
"max_num_seqs": 4,
"max_num_seqs": 16,
"gpu_memory_utilization": 0.95,
"enforce_eager": false,
"enable_prefix_caching": false,
"max_num_batched_tokens": 32768,
"devices": "0,1,2,3",
"tensor_parallel_size": 4,
"devices": "0,1",
"tensor_parallel_size": 2,
"hf_overrides": {
"rope_parameters": {
"mrope_section": [0, 32, 32],
@@ -65,13 +65,14 @@
"stage_id": 1,
"max_num_seqs": 1,
"enforce_eager": true,
"devices": "4,5,6,7",
"devices": "2,3",
"distributed_executor_backend": "mp",
"omni_kv_config": {
"need_recv_cache": true
},
"parallel_config": {
"tensor_parallel_size": 4,
"tensor_parallel_size": 2,
"vae_patch_parallel_size": 2,
"enable_expert_parallel": true
},
"input_connectors": {
@@ -114,8 +115,8 @@
},
"task": "ti2i",
"bot-task": "think_recaption",
"max-concurrency": 4,
"num-prompts": 16,
"max-concurrency": 16,
"num-prompts": 96,
"skip-performance-assertion": true,
"baseline": {
"throughput_qps": 0.1,
vllm_omni/diffusion/models/hunyuan_image3/pipeline_hunyuan_image3.py
+4 -1
View File
@@ -865,7 +865,9 @@ class HunyuanImage3Pipeline(
output, sections = out["output"], out["sections"]
# 4. Encode conditional images
if batch_cond_image_info is not None and len(batch_cond_image_info[0]) > 0:
# Skip encoding if AR KV reuse is enabled
has_ar_kv = kwargs.get("ar_kv_data")
if batch_cond_image_info is not None and len(batch_cond_image_info[0]) > 0 and not has_ar_kv:
cond_vae_images, cond_timestep, cond_vit_images = self._encode_cond_image(
batch_cond_image_info, cfg_factor[mode]
)
@@ -1459,6 +1461,7 @@ class HunyuanImage3Pipeline(
num_inference_steps=num_inference_steps,
guidance_scale=guidance_scale,
batch_cond_image_info=batch_cond_image_info,
**ar_kv_kwargs,
)
model_inputs.update(ar_kv_kwargs)