Fairplace
arrow_back Retour aux freelances
F

AI Video Engineer for ROLLCALL Backend

Freelancer

Partager:
placeFR home_workTélétravail assignmentCDD publicOffre agrégée · FR

eventPubliée le 24 août 2026 · verifiedNous avons confirmé le 30 août 2026 qu'elle est toujours active

US$ 30 – US$ 250 par projet

À propos de l'offre

Senior AI Video / GPU Engineer Needed – Wan2.2 + RunPod A100 + PyTorch/CUDA I need an experienced AI/GPU engineer to finish and productionize an existing AI video-generation backend for a platform called ROLLCALL. This is NOT a website design job. The website is already built. I need someone who specializes in GPU inference, Python, PyTorch/CUDA environments, AI video models, and production API deployment. CURRENT SYSTEM We already have: RunPod NVIDIA A100-SXM4 80GB GPU Wan2.2 I2V A14B Approximately 118GB of Wan2.2 model files already downloaded Persistent /workspace storage Python PyTorch/CUDA FastAPI/Uvicorn worker FFmpeg Existing website integration Existing REST API running on port 3010 Current API routes include: GET /v1/health POST /v1/generate GET /v1/jobs/{job_id} GET /v1/files/{filename} The website can communicate with RunPod successfully. GPU health verification works. The generation API successfully accepts a request, creates a real job_id, returns that job ID to the website, and allows the website to poll the job status. CURRENT TECHNICAL PROBLEM Wan2.2 is currently failing during Python startup because of a FlashAttention/PyTorch ABI compatibility problem. Current error includes: flash_attn_2_cuda...so: undefined symbol... I do NOT want someone randomly reinstalling packages until something works. I want an engineer who understands PyTorch, CUDA, NVIDIA GPUs, FlashAttention, compiled Python/CUDA extensions, ABI compatibility, and production AI inference environments. YOUR JOB First audit the complete existing environment before changing anything. Determine the correct compatible combination of: NVIDIA driver CUDA Python PyTorch torchvision/torchaudio where applicable FlashAttention Wan2.2 dependencies Then repair or rebuild the Python runtime cleanly if necessary. You must also: Preserve the existing ~118GB Wan2.2 model files. Preserve existing ROLLCALL project files and generated videos. Verify Wan2.2 imports successfully. Verify generate.py starts successfully. Verify T5, VAE and Wan2.2 model checkpoints. Verify A100 CUDA inference. Verify FFmpeg. Review and stabilize the FastAPI worker. Verify generation requests immediately return a valid job ID. Verify job status can be polled reliably. Implement reliable queued, generating, processing/stitching, completed and failed states. Return useful error information when a generation fails. Ensure completed MP4 files are accessible to the website. Verify image-to-video generation using Wan2.2 I2V A14B. Test the entire system from the ROLLCALL website through RunPod and back. Make the worker start/recover correctly after a pod/container restart. Create a reproducible startup/deployment configuration so we do not have to manually repair the environment again. Document the exact final Python/PyTorch/CUDA/FlashAttention/package versions. IMPORTANT Do NOT delete, replace or redownload the approximately 118GB Wan2.2 model unless there is a genuine technical reason and I approve it first. Do NOT consider the project finished simply because the health endpoint returns HTTP 200. DEFINITION OF DONE I will consider this project complete only when this complete workflow works: ROLLCALL Website → RunPod API → Job Created → Wan2.2 I2V A14B → NVIDIA A100 GPU Generation → MP4 Created → Job Completed → Video Returned → Video Plays Correctly on ROLLCALL Website A real end-to-end video generation must successfully complete before final acceptance. I also want the completed environment documented and reproducible so a future RunPod restart does not require rebuilding everything manually. REQUIRED EXPERIENCE Please apply only if you have strong experience with several of the following: Python, PyTorch, CUDA, NVIDIA A100/H100, RunPod, Wan2.1/Wan2.2, FlashAttention, Hugging Face, diffusion/video models, FastAPI, REST APIs, Linux, FFmpeg, Docker and production GPU inference. Experience deploying large AI video models such as Wan2.x, HunyuanVideo, CogVideoX, Stable Video Diffusion or similar models is strongly preferred. IMPORTANT SCREENING QUESTION Start your proposal with: A100-WAN22 Then answer this question: If flash_attn_2_cuda.so produces an undefined symbol error when importing Wan2.2, what would you inspect before reinstalling FlashAttention? Explain specifically how you would verify PyTorch version, CUDA compatibility, NVIDIA driver, Python version, FlashAttention build/wheel compatibility and ABI compatibility. Do not simply answer “I will reinstall FlashAttention.” Generic proposals will not be considered. I am looking for someone who can inspect the entire stack, identify the root cause, make the correct repair once, and deliver a production-ready system.

Continuez à lire gratuitement

Créez un compte gratuit pour voir l'offre complète et postuler.

  • badgePortfolio visible par les entreprises
  • notificationsAlerte de nouvelle offre par e-mail
  • favoriteToujours gratuit, sans piège
Créer un compte gratuit J'ai déjà un compte

Offres similaires

Terjemahan Inggris ke Indonesia

Freelancer

placeID 31 août

Listing Produk Pakaian Online

Freelancer

placeID 31 août

We're expanding SkillToDollar into the MENA freelancer market.

Freelancer

placePH 31 août

Instagram Account Management

Freelancer

placeJP 31 août

Rust Proxy Speed Optimization

Freelancer

placeNL 31 août

Residential Remodel Drafting Plans

Freelancer

placeES 31 août