Managed Compute Qwen3-Coder: /openai/v1/responses returns zero token usage while /chat/completions reports correct usage

Hello,

We are observing incorrect token usage reporting from the OpenAI-compatible

Responses API for a Qwen3-Coder model deployed using Azure AI Foundry

Managed Compute.

The inference itself succeeds and returns the expected generated text, but

all usage counters in the Responses API result are zero.

The same deployment reports correct token usage through Chat Completions.

Environment

Managed Compute deployment:

qwen3-coder-30b-a3b-fp8

Model asset:

azureml://registries/azure-huggingface/models/qwen--qwen3-coder-30b-a3b-instruct-fp8/versions/3

Deployment template:

azureml://registries/azure-huggingface/deploymenttemplates/qwen--qwen3-coder-30b-a3b-instruct-fp8--256k-nvidia-h100/labels/latest

Accelerator:

1 x H100 80 GB

Runtime fingerprint returned by Chat Completions:

vllm-0.22.0-ep-e7db14b8

Reproduction 1: Chat Completions

curl -sS -X POST \

"https://<REDACTED>.cognitiveservices.azure.com/openai/deployments/qwen3-coder-30b-a3b-fp8/chat/completions?api-version=2025-04-01-preview" \

-H "Content-Type: application/json" \

-H "api-key: <REDACTED>" \

-d '{

"messages": [

  {

    "role": "user",

    "content": "Reply only: OK"

  }

],

"max_tokens": 32,

"temperature": 0,

"stream": false
```  }'

The response is successful and reports correct usage:

{

  "model": "qwen--qwen3-coder-30b-a3b-instruct-fp8",

  "choices": [

```scala
{

  "message": {

    "role": "assistant",

    "content": "OK"

  }

}
```  ],

  "system_fingerprint": "vllm-0.22.0-ep-e7db14b8",

  "usage": {

```yaml
"prompt_tokens": 12,

"completion_tokens": 2,

"total_tokens": 14
```  }

}

Reproduction 2: Responses API

-----------------------------

curl -sS -X POST \

  "https://<REDACTED>.cognitiveservices.azure.com/openai/v1/responses" \

  -H "Content-Type: application/json" \

  -H "api-key: <REDACTED>" \

  -d '{

```python
"model": "qwen3-coder-30b-a3b-fp8",

"input": "Reply only: OK",

"max_output_tokens": 32,

"temperature": 0,

"stream": false
```  }'

The response is also successful and contains the expected text "OK".

However, usage is:

{

  "input_tokens": 0,

  "output_tokens": 0,

  "total_tokens": 0,

  "input_tokens_details": {

"cached_tokens": 0

  "output_tokens_details": {

```yaml
"reasoning_tokens": 0
```  }

}

Streaming behavior

------------------

The same discrepancy occurs with streaming:

- Chat Completions with stream=true and

  stream_options.include_usage=true reports 12 input tokens,

  2 output tokens and 14 total tokens.

- The final response.completed event from /openai/v1/responses reports

  input_tokens=0, output_tokens=0 and total_tokens=0.

Expected behavior

-----------------

The Responses API should report the actual input and output token counts,

consistent with the Chat Completions API for the same deployment, prompt

and generated output.

The zero values originate from the Foundry

Managed Compute serving path or its packaged vLLM runtime.

We found vLLM PR #22667, which fixed zero Responses API usage for the

GPT-OSS Harmony context. However, this deployment uses Qwen3-Coder and

appears to follow the non-Harmony Responses path, so that fix may not

apply here.

Questions

---------

1. Can you reproduce the zero usage values for this Managed Compute

   deployment and model?

2. Is /openai/v1/responses processed by a Foundry compatibility layer,

   or passed directly to the Microsoft-curated vLLM runtime?

3. Is there a newer deployment template or runtime image that fixes

   Responses API token usage for Qwen models?

1. If the issue is in upstream vLLM, can you confirm this and provide

   the relevant upstream issue or route it to the appropriate team?
   
Thank you.