Skip to content

[Proposal] Add RunPod Serverless Support for DeepSeek Endpoint Overrides #238

Description

@joeblew999

🚀 The Problem / Use Case

We want to run a private, cost-effective deployment of DeepSeek-R1 (Distilled 8B/32B) using RunPod Serverless instead of paying for the official DeepSeek API or a flat-rate dedicated cloud GPU.While the deepseek-v4-for-copilot extension supports custom baseUrl settings, connecting it to a scale-to-zero serverless endpoint (like RunPod) introduces a cold start latency (20–40 seconds) on the very first API request. This causes the initial connection to occasionally hang or fail if standard HTTP timeouts are strictly enforced.

💡 Proposed Solution

Formally document and optimize the configuration pipeline to support RunPod Serverless as a backend provider via standard OpenAI-compatible endpoints.Base URL Mapping: Override "deepseek-copilot.baseUrl" to route directly to https://runpod.ai{endpoint_id}/vllm/v1.Model Identifiers: Map "deepseek-copilot.modelIdOverrides" to the exact Hugging Face identifiers deployed on the RunPod worker (e.g., deepseek-ai/DeepSeek-R1-Distill-Llama-8B).

Timeout Resilience (Optional Feature Request)

If applicable, expose an internal request timeout configuration setting (e.g., deepseek-copilot.requestTimeout) or increase the default network timeout to 60 seconds to prevent cold-start failures.

Proposed Configuration Changes:

(settings.json)json{
"deepseek-copilot.baseUrl": "https://runpod.ai",
"deepseek-copilot.modelIdOverrides": {
"deepseek-v4-flash": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B",
"deepseek-v4-pro": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B"
}
}
Use code with caution.

Alternatives :

ConsideredDedicated RunPod Instances: Fast execution and no cold starts, but costs run 24/7 ($150+ / month) even when idle.Ollama (Local Deployment): Free, but restricted by local hardware (VRAM limits) and slows down local IDE compilation performance.

BenefitsCost Efficiency:

Zero cost when the developer is not writing code or using Copilot Chat.Privacy: Model weights and execution run completely inside an isolated, private RunPod container.Native Integration: Retains all of Copilot's superior UI features, agent tools, and file workspace searching.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions