DeepSeek FIM API 原生可用，但 Continue 的 autocomplete/FIM stream 链路报错

Before submitting your bug report

• I've tried using the "Ask AI" feature on the Continue docs site to see if the docs have an answer
• I'm not able to find a related conversation on GitHub discussions that reports the same bug
• I'm not able to find an open issue that reports the same bug
• I've seen the troubleshooting guide on the Continue Docs

Relevant environment info

* OS: Windows
* Continue version: v1.2.22; also tested an older version from around 8 months ago
* IDE version: VS Code on Windows
* Model: deepseek-v4-pro
* Provider: deepseek
* Role: autocomplete
* File type tested: Markdown (`.md`)
* Config:


models:
  - name: DeepSeek Autocomplete
    provider: deepseek
    model: deepseek-v4-pro
    apiBase: https://api.deepseek.com/beta
    roles:
      - autocomplete
    defaultCompletionOptions:
      temperature: 0.0
      maxTokens: 64
    autocompleteOptions:
      disable: false
      maxPromptTokens: 1024
      debounceDelay: 800
      modelTimeout: 10000
      maxSuffixPercentage: 0.2
      prefixPercentage: 0.6
      onlyMyCode: false


API key is omitted/redacted.

Description

DeepSeek FIM autocomplete fails in Continue, while the native DeepSeek FIM API works correctly.

Continue does trigger the FIM autocomplete path for Markdown files, but autocomplete fails with an internal error from llm_stream_fim.

The same DeepSeek FIM endpoint works when called directly with PowerShell:

• Endpoint: https://api.deepseek.com/beta/completions
• Model: deepseek-v4-pro
• stream: true
• Result: SSE chunks are returned successfully, ending with [DONE]

So the issue does not seem to be the DeepSeek API key, network, endpoint, model name, or Markdown triggering. It appears to be related to Continue's DeepSeek FIM stream handling.

To reproduce

1. Configure a DeepSeek autocomplete model in Continue:

- name: DeepSeek Autocomplete
  provider: deepseek
  model: deepseek-v4-pro
  apiBase: https://api.deepseek.com/beta
  roles:
    - autocomplete
  defaultCompletionOptions:
    temperature: 0.0
    maxTokens: 64
  autocompleteOptions:
    disable: false
    maxPromptTokens: 1024
    debounceDelay: 800
    modelTimeout: 10000
    maxSuffixPercentage: 0.2
    prefixPercentage: 0.6
    onlyMyCode: false

2. Enable Continue autocomplete.

3. Select DeepSeek Autocomplete as the current autocomplete model.

4. Enable FIM autocomplete over Next Edit.

5. Open a Markdown file.

6. Type something like:

## Fourier series

Let \(f(x)\) be defined on

7. Wait for autocomplete.

8. Continue triggers llm_stream_fim, but no autocomplete suggestion appears and the extension logs an error.

Log output

[Extension Host] [@continuedev] error: {"context":"llm_stream_fim","model":"deepseek-v4-pro","provider":"deepseek","useOpenAIAdapter":true}

Error generating autocompletion: Error:
    at streamResponse2 (...\continue.continue-1.2.22-win32-x64\out\extension.js:178977:11)
    at async streamSse2 (...\continue.continue-1.2.22-win32-x64\out\extension.js:179048:20)
    at async DeepSeekApi.fimStream (...\continue.continue-1.2.22-win32-x64\out\extension.js:243838:26)
    at async Deepseek.streamFim (...\continue.continue-1.2.22-win32-x64\out\extension.js:318216:30)



Native DeepSeek FIM API test succeeds:

PowerShell request:

$body = @{
  model = "deepseek-v4-pro"
  prompt = "## Fourier series`n`nLet \(f(x)\) be defined on "
  suffix = " and be piecewise smooth"
  max_tokens = 64
  temperature = 0
  stream = $true
} | ConvertTo-Json

Invoke-RestMethod `
  -Uri "https://api.deepseek.com/beta/completions" `
  -Method Post `
  -Headers @{
    "Authorization" = "Bearer <redacted>"
    "Content-Type" = "application/json"
  } `
  -Body $body

The response returns SSE data chunks and ends with:

data: [DONE]

Start by tracing the llm_stream_fim path through DeepSeekApi.fimStream and Deepseek.streamFim; the report only identifies the bundled extension.js locations. Compare Continue's handling with the successful DeepSeek SSE response and reproduce the Markdown autocomplete case. Done means the configured DeepSeek model produces an autocomplete suggestion without the stream error.