[Bug] zeroshot infer error

这是日志
2026-09-08 21:14:58,492 - wall_x._vendor.harrix.serving.scheduler - INFO - Batch processed in 89.4ms, throughput: 11.2 req/s
2026-09-08 21:14:58,496 - wall_x._vendor.harrix.serving.websocket_policy_server - INFO - Connection from ('192.168.5.6', 57346) closed
2026-09-08 21:14:59,572 - websockets.server - INFO - connection open
2026-09-08 21:14:59,572 - wall_x._vendor.harrix.serving.websocket_policy_server - INFO - Connection from ('192.168.5.6', 57354) opened
2026-09-08 21:14:59,620 - wall_x._vendor.harrix.serving.scheduler - INFO - Processing batch of 1 requests
2026-09-08 21:14:59,621 - wall_x._vendor.harrix.serving.policy.wall_x_policy - INFO - WallXPolicy.infer_batch: processing 1 observations
[21:14:59] [ ROBOT ] [ROBOT.Robot] INFO EX001RobotPreprocessor: using 'camera_front' as 'face_view'
[21:14:59] [ ROBOT ] [ROBOT.Robot] INFO EX001RobotPreprocessor: using 'camera_left' as 'left_wrist_view'
[21:14:59] [ ROBOT ] [ROBOT.Robot] INFO EX001RobotPreprocessor: using 'camera_right' as 'right_wrist_view'
2026-09-08 21:14:59,657 - wall_x.utils.timers - INFO - get_text_for_action took 0.010 ms to execute
2026-09-08 21:14:59,659 - wall_x.utils.timers - INFO - resize_images took 2.147 ms to execute
2026-09-08 21:14:59,661 - wall_x.utils.timers - INFO - preprocesser_call took 2.385 ms to execute
2026-09-08 21:14:59,662 - wall_x.utils.timers - INFO - convert_action_token_id and post took 0.188 ms to execute
2026-09-08 21:14:59,662 - wall_x.utils.timers - INFO - construct_model_input took 5.187 ms to execute
2026-09-08 21:14:59,707 - wall_x.utils.timers - INFO - generate_flow_action_batch took 45.366 ms to execute
2026-09-08 21:14:59,707 - wall_x._vendor.harrix.serving.policy.wall_x_policy - ERROR - Batch inference failed: shape mismatch: value tensor of shape [32, 2048] cannot be broadcast to indexing result of shape [0, 2048]
Traceback (most recent call last):
File "/home/user/wall-x/wall_x/_vendor/harrix/serving/policy/wall_x_policy.py", line 296, in infer_batch
model_outputs = self.model_wrapper.infer_flow_action_batch(
File "/home/user/wall-x/wall_x/_vendor/harrix/serving/_wallx_infer/model_wrapper.py", line 639, in infer_flow_action_batch
model_output = self.model.generate_flow_action(
File "/home/user/anaconda3/envs/wallx/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 124, in decorate_context
return func(*args, **kwargs)
File "/home/user/wall-x/wall_x/model/qact/qwen2_5/modeling_qwen2_5_vl_act.py", line 1806, in generate_flow_action
ctx = self._prepare_flow_action_inputs(
File "/home/user/wall-x/wall_x/model/qact/qwen2_5/modeling_qwen2_5_vl_act.py", line 2078, in _prepare_flow_action_inputs
inputs_embeds[flow_action_mask] = action_embed
RuntimeError: shape mismatch: value tensor of shape [32, 2048] cannot be broadcast to indexing result of shape [0, 2048]

部署和推理代码

nohup bash scripts/run_serving.sh --checkpoint-path /home/user/models/wall-oss-0.5 --port 32195 --serialize-actions --max-batch-size 2 --robot-type ex001 --debug > wallx_serve_09082.log 2>&1 &

import asyncio
import base64
import json
import msgpack
import websockets

WS_URI = "ws://192.168.5.80:32195"
INSTRUCTION = "pick the cup"

dummy_state = {
    "follow1_pos": [0.0] * 7,
    "follow2_pos": [0.0] * 7,
}

def img_to_b64(img_path):
    with open(img_path, "rb") as f:
        return base64.b64encode(f.read()).decode("utf-8")

dummy_views = {
    "camera_front": img_to_b64("/home/zkl/lerobot/sdk_robot/examples/HeadRGBStream_screenshot_08.09.2026.png"),
    "camera_left": img_to_b64("/home/zkl/lerobot/sdk_robot/examples/LeftArmStream_screenshot_08.09.2026.png"),
    "camera_right": img_to_b64("/home/zkl/lerobot/sdk_robot/examples/RightArmStream_screenshot_08.09.2026.png"),
}

payload = {
    "state": dummy_state,
    "views": dummy_views,
    "instruction": INSTRUCTION,
}

async def main():
    async with websockets.connect(WS_URI) as ws:
        meta_raw = await ws.recv()
        meta = msgpack.unpackb(meta_raw, raw=False)
        print("收到 metadata:", json.dumps(meta, indent=2, ensure_ascii=False))

        await ws.send(msgpack.packb(payload, use_bin_type=True))

        resp_raw = await ws.recv()
        if isinstance(resp_raw, str):
            print("服务端文本错误：")
            print(resp_raw)
            return

        resp = msgpack.unpackb(resp_raw, raw=False)
        print("模型输出 action:")
        print(json.dumps(resp, indent=2, ensure_ascii=False))

        print("左臂:", resp.get("follow1_pos"))
        print("右臂:", resp.get("follow2_pos"))

if __name__ == "__main__":
    asyncio.run(main())

Navigation Menu

Repository navigation

[Bug] zeroshot infer error #120

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

[Bug] zeroshot infer error #120

Description

Activity

zkailinzhang commented on Sep 8, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions