Revert gradio 5.x bump (HF Dockerfile forces 4.44); pin starlette+fastapi instead

requirements.txt

@@ -1,24 +1,38 @@
-# Gradio 5.x for ChatInterface in messages mode + Spaces ZeroGPU support.
+# NOTE for the next person debugging this file:
 #
-# Why 5.x: Gradio 4.44 calls Starlette's TemplateResponse with the old
-# (name, context) signature. Starlette ≥ 0.37 (which fastapi >= 0.111
-# transitively pins) expects (request, name, context) — so the request
-# dict ends up being used as the template name, blowing up jinja2's
-# cache lookup with "TypeError: unhashable type: 'dict'" on every
-# request. Pinning jinja2 didn't help (the bug is upstream of jinja2).
-# Gradio 5.x adopted the new Starlette signature and fixes this.
-gradio>=5.0,<6
+# HF Spaces ZeroGPU has a Dockerfile that INJECTS these versions on
+# top of whatever's in this file:
+#
+#   gradio[oauth]==4.44.0
+#   spaces==0.50.2
+#   torch<=2.11.0
+#   uvicorn>=0.14.0
+#
+# You cannot upgrade gradio past 4.44 on a ZeroGPU Space — pip will
+# fail dependency resolution. Don't try (5467dbe / df84d94 already did
+# and burned 30 min of rebuilds).
+#
+# Gradio 4.44 calls Starlette's TemplateResponse with the old
+# positional signature  (name, context). Starlette ≥ 0.38 raises
+# `TypeError: unhashable type: 'dict'` at jinja2 cache lookup because
+# it interprets the request as the template name. Pin starlette + the
+# matching fastapi to the last versions that still tolerate the old
+# signature.
+
+# Pinned web stack that's known-compatible with Gradio 4.44.0:
+fastapi>=0.111,<0.115
+starlette>=0.37.2,<0.40
+
+# Spaces SDK (the @spaces.GPU decorator). HF pins this too but
+# declaring explicitly so we know the version we tested against.
 spaces
 
-# Model loading + inference. Pinning major versions so a transformers
-# breaking change on HF can't silently nuke the Space at next rebuild.
+# Model loading + inference.
 transformers>=4.44,<5
-torch>=2.3,<3
 accelerate>=0.30
 sentencepiece
 protobuf
 
-# Quantization not strictly needed on the A100 (BioMistral-7B fits in
-# bf16 with 26 GB free) but kept in case we ever swap to a larger model
-# or want to test bnb 4-bit memory wins locally.
+# bitsandbytes is optional (BioMistral-7B fits in bf16 on A100 with
+# room to spare). Kept for future swap to a larger model.
 bitsandbytes>=0.43