Align safetensors with Transformers 5.15

docs/LEARNINGS.md

@@ -2,6 +2,7 @@
 
 - A painless Hugging Face model-ID input still needs an immutable deployment boundary. Resolve the requested repository to an exact commit on CPU, derive a content-addressed schema-v2 profile from model metadata plus a versioned compatibility policy, and reject unknown architectures or quantization before downloading tens of gigabytes. Qwen3.5 block-FP8 and ModelOpt NVFP4 can share the current vLLM adapter; Muse-Glimmer needs the released Transformers 5.15 implementation.
 - vLLM 0.27.1 pins Torch 2.13.0, so the B300 CUDA 13.2 image must move as a unit to Torch 2.13.0, torchvision 0.28.0, Transformers 5.15.0, and vLLM 0.27.1. Keep vLLM off older CUDA 12.8 images whose tested Torch stack is 2.10.0. For ComfyUI co-residency, pass `kv_cache_memory_bytes` explicitly (currently 12 GiB) and start in eager mode; the default GPU utilization fraction is inappropriate on a B300 shared with image/video models.
+- Transformers 5.15.0 raises its safetensors floor to 0.8.0. Keeping the older 0.7.0 ComfyUI pin makes Modal's image dependency resolver fail before deployment, so the backend upgrade must include safetensors 0.8.0 in the same runtime fingerprint.
 
 ## 2026-08-18
 

runtime_environment.py

@@ -51,7 +51,7 @@ _REMOTE_RUNTIME_PACKAGES = (
     "pyopengl==3.1.10",
     "pyyaml==6.0.3",
     "requests==2.32.5",
-    "safetensors==0.7.0",
+    "safetensors==0.8.0",
     "scipy==1.17.0",
     "sentencepiece==0.2.1",
     "simpleeval==1.0.7",

tests/test_runtime_environment.py

@@ -75,6 +75,7 @@ def test_remote_environment_is_fully_pinned(runtime_environment_module: Any) ->
     assert "comfy-kitchen==0.2.31" in runtime_packages
     assert "pyopengl==3.1.10" in runtime_packages
     assert "pypdf==6.16.1" in runtime_packages
+    assert "safetensors==0.8.0" in runtime_packages
     assert "huggingface-hub==1.28.0" in runtime_packages
     assert "transformers==5.15.0" in runtime_packages
     assert "simpleeval==1.0.7" in runtime_packages