Track 1ROCm + vLLMQwen3-Coder-NextCUDA to ROCm

ROCmPilot

Multi-agent ROCm migration cockpit for PyTorch and vLLM workloads.

Agent run
Ready to audit the selected workload
idle
Select a workload and start the audit.
Agent War Room
Lead-led discussion with replies, objections, and shared memory the agents reuse later.
standbylead: No lead assigned
Repo Doctor waiting
Migration Planner waiting
Build Runner waiting
Benchmark Agent waiting
Report Agent waiting
Orchestrator waiting
CUDA/ROCm Coach Agent
Separate helper agent for developer questions, using the current findings and memory.
sidecar agent

CUDA/ROCm Coach

I am the CUDA/ROCm Coach Agent. Ask me what a finding means, how ROCm differs from CUDA, or how to validate PyTorch/vLLM on AMD.

Context: 4 findings, 3 patches, 0 memories.

Migration Kit Agent
Builds a downloadable implementation file from repo findings, patches, and shared memory.
export agent

Target

Qwen vLLM CUDA Starter

4

findings

3

files

0

memories

Package preview

  • Generates qwen-vllm-cuda-starter-migration-kit.md
  • Includes patch previews for src/inference/device.py, Dockerfile.rocm, scripts/serve-rocm.sh
  • Adds a ROCm validation script for PyTorch HIP detection and vLLM serving
  • References shared memory from the agent war room
Migration findings
CUDA assumptions, ROCm blockers, and recommended fixes.
SeverityCategoryLocationFix
Findings appear after the Repo Doctor stage starts.
Evidence panels
Patch previews, terminal logs, and final report output.
Patch previews unlock after the Migration Planner stage.