GPU pending issue when draft-target is deployed using triton server

When I deploy the model with triton server, it looks like this：

I0408 07:45:34.906138 205499 server.cc:674] 
+----------------+---------+--------+
| Model          | Version | Status |
+----------------+---------+--------+
| llama_3_2_1B  | 1       | READY  |
| llama_3_3_70B  | 1       | READY  |
| ensemble       | 1       | READY  |
| postprocessing | 1       | READY  |
| preprocessing  | 1       | READY  |
+----------------+---------+--------+

Target model: llama-3.3-70B-instruct
Draft model : llama-3.2-1B-instruct

• case 1: use grpc client request one prompt Successful✅

[TensorRT-LLM][WARNING] Trying to remove block 9 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 9 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 9 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 9 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 9 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 9 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 9 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 9 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 10 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 10 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 10 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 10 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 10 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 10 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 10 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 10 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 170 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 170 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 170 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 170 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 170 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 170 by 0 that is not in hash map

• case 2: I request a prompt through the grpc client. When inference is in progress, I send another request to simulate IFB, and the order of multiple online requests in the time dimension is 100%. At this time, the GPU utilization rate is 100% and there is no fluctuation. The TX RX data volume is relatively large. Inference does not seem to continue, and there is no error message. It is stuck. Failed❌

[TensorRT-LLM][WARNING] Trying to remove block 9 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 9 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 9 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 9 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 9 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 9 by 0 that is not in hash map
# It will get stuck here and no log will be generated

and GPUs status:

 Device 0 [NVIDIA H20] PCIe GEN 5@16x RX: 364.4 MiB/s TX: 90.59 MiB/s
 GPU 1980MHz MEM 2619MHz TEMP  34°C FAN N/A% POW 115 / 500 W
 GPU[|||||||||||||||||||||||||100%] MEM[|||||||||||71.022Gi/95.577Gi]

 Device 1 [NVIDIA H20] PCIe GEN 5@16x RX: 1.309 GiB/s TX: 223.1 MiB/s
 GPU 1980MHz MEM 2619MHz TEMP  31°C FAN N/A% POW 124 / 500 W
 GPU[|||||||||||||||||||||||||100%] MEM[|||||||||||71.068Gi/95.577Gi]

 Device 2 [NVIDIA H20] PCIe GEN 5@16x RX: 5.047 GiB/s TX: 602.4 MiB/s
 GPU 1980MHz MEM 2619MHz TEMP  33°C FAN N/A% POW 123 / 500 W
 GPU[|||||||||||||||||||||||||100%] MEM[|||||||||||71.068Gi/95.577Gi]

 Device 3 [NVIDIA H20] PCIe GEN 5@16x RX: 530.0 MiB/s TX: 150.1 MiB/s
 GPU 1980MHz MEM 2619MHz TEMP  31°C FAN N/A% POW 126 / 500 W
 GPU[|||||||||||||||||||||||||100%] MEM[|||||||||||71.068Gi/95.577Gi]

 Device 4 [NVIDIA H20] PCIe GEN 5@16x RX: 9.167 GiB/s TX: 1.097 GiB/s
 GPU 1980MHz MEM 2619MHz TEMP  34°C FAN N/A% POW 127 / 500 W
 GPU[|||||||||||||||||||||||||100%] MEM[|||||||||||71.068Gi/95.577Gi]

 Device 5 [NVIDIA H20] PCIe GEN 5@16x RX: 4.246 GiB/s TX: 512.2 MiB/s
 GPU 1980MHz MEM 2619MHz TEMP  31°C FAN N/A% POW 125 / 500 W
 GPU[|||||||||||||||||||||||||100%] MEM[|||||||||||71.068Gi/95.577Gi]

 Device 6 [NVIDIA H20] PCIe GEN 5@16x RX: 9.461 GiB/s TX: 1.027 GiB/s
 GPU 1980MHz MEM 2619MHz TEMP  34°C FAN N/A% POW 127 / 500 W
 GPU[|||||||||||||||||||||||||100%] MEM[|||||||||||71.068Gi/95.577Gi]

 Device 7 [NVIDIA H20] PCIe GEN 5@16x RX: 799.7 MiB/s TX: 173.2 MiB/s
 GPU 1980MHz MEM 2619MHz TEMP  31°C FAN N/A% POW 123 / 500 W
 GPU[|||||||||||||||||||||||||100%] MEM[|||||||||||70.834Gi/95.577Gi]

Anyone got a solution?

Start by reproducing the deployment with the Triton server and gRPC client using the two request patterns described: one request, then overlapping requests during inference. Compare the logs and GPU behavior, focusing on the repeated block-removal warnings and the point where output stops. Done means concurrent requests complete without hanging and the warnings are resolved or explained.