β‘ ComfyUI Colab Ultra-Flash Assets (ramusgill/comfyui-colab-flash)
Pre-built, zero-bottleneck runtime archives, pre-compiled CUDA 13 / Blackwell kernels, custom nodes, and LoRA manifests engineered for sub-second API latency and ultra-fast multimodal Video+Audio inference (MiniMax-H3 Turbo + Qwen-Image-Edit 2511) on Google Colab GPUs.
π¦ Repository Structure
| Path | Size | Description |
|---|---|---|
h3_env_py313_cu130_flash.tar.zst |
1.46 GiB | Clean Python 3.13 + ComfyUI 0.39.0 core (--system-site-packages on CUDA 13.0, zero duplicate nvidia/*/torch bloat, onnxruntime-gpu 1.31.1 CUDA 13, FP8 TensorCore dynamic scale="recalculate" & RoPE patches pre-applied). |
custom_nodes.tar.zst |
62.7 MiB | 84 curated, sanitized custom nodes (9.4x slimmer than raw checkout; ReActor synchronous boot-download disabled; broken 3D/LLM stalls excluded). |
loras_metadata.tar.zst |
8.7 KiB | Pre-computed SHA256 .metadata.json index for all 87 LoRAs β eliminates 27 GB of startup disk hashing in comfyui-lora-manager. |
triton_cache_cu130_blackwell.tar.zst |
112.8 KiB | 49 pre-compiled Blackwell (sm_120, CUDA 13.0) Triton PTX/cubin kernels (/root/.triton/cache). |
wheels/ |
1.68 MiB | sageattn3-1.0.0-cp313-linux_x86_64.whl (Blackwell FP4) + sageattention-2.2.0-py3-none-any.whl (SageAttn3 FP4 / ComfyKitchen INT8 / SDPA dispatcher). |
loras/ + loras_manifest.json |
87 files | Personal & SOTA Turbo LoRAs (minimax_h3_ref2v_turbo_4step_v0.1_comfyui_resized_avg_rank_21_bf16, minimax_h3_pdmd_2step_lora_avg_rank_38_bf16, Qwen-Edit Lightning, character/style LoRAs). |
text_encoders/ & vae/ |
9.19 GiB | qwen_2.5_vl_7b_fp8_scaled.safetensors + qwen_image_vae.safetensors mirrored for multi-connection aria2c. |
MiniMax_H3_Turbo_Colab_Flash.ipynb |
48 KiB | Ready-to-run Colab Flash notebook (2-tier aria2c, --fast-disk, --use-ck-attention, /dev/shm RAM-disk, QUIC tunnel, GPU auto-warmup & watchdog). |
bootstrap_remote.py |
32 KiB | Headless remote bootstrapper for colab CLI automation. |
π Verified Live Benchmarks (Colab G4 β RTX PRO 6000 Blackwell 95 GiB, CUDA 13.0)
| Metric | Legacy Setup | Ultra-Flash Optimized | Speedup |
|---|---|---|---|
| Environment + Nodes Download | 5.46 GiB + 587 MiB | 1.46 GiB + 62.7 MiB | 4.0x smaller |
| ComfyUI Boot (84 Custom Nodes) | 52.3s (with import errors) | 14.7s (0 import errors) | 3.55x faster |
| Cold 1st Generation (37.6 GiB to VRAM) | 140.9s (fast_disk=False) |
14.0s (--fast-disk + caches) |
10.0x faster |
Warm 56f (2.33s Video+Audio 640x640 H.264) |
~9.5s | 3.63s (4-step Rank-21) / 2.43s (2-step PDMD) | Up to 3.9x faster |
Warm 124f (5.17s Video+Audio 640x640 H.264) |
~24.0s | 11.75s (INT8) / 9.73s (W4A8 14GB) / 6.06s (2-step) | Up to 4.0x faster |
π Zero-Secret Tunnel Configuration
Store CLOUDFLARE_TOKEN in Google Colab Secrets (userdata.get('CLOUDFLARE_TOKEN')) or pass it as an environment variable when running bootstrap_remote.py. If omitted, the notebook automatically falls back to an ephemeral trycloudflare.com QUIC tunnel.
- Downloads last month
- 57,871
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support