Move nemo-toolkit install to runtime to unblock gradio 6.20 build

The previous upgrade failed at build time because:
- gradio 6.20 requires huggingface-hub>=1.2
- nemo-toolkit[tts]==2.4.0 pins transformers<=4.52 (4.51.x)
- transformers 4.51.x requires huggingface-hub<1.0

pip refuses to resolve these together. Move the nemo install to
create_env.py so it happens after we've pinned transformers==5.3.0
+ hub>=1.2 with --no-deps — sidestepping the resolver conflict.

- requirements.txt: remove nemo-toolkit[tts]; list its common runtime
deps explicitly (librosa, soundfile, omegaconf, webdataset, pandas,
etc.) since --no-deps skips them.
- create_env.py: rewrite as a 3-step install — pin transformers+hub,
install nemo with --no-deps, cap numpy<2.0.
- README.md: document the new install dance.
- app.py: update the startup-order docstring.

Co-Authored-By: Claude <noreply@anthropic.com>

README.md

@@ -48,5 +48,15 @@ generation parameters — re-point the Space without touching code.
 - The model repo is private: set the `HF_TOKEN` Space secret.
 - ZeroGPU: the model is loaded once at startup; each request only runs
   generation inside the GPU context.
-- `create_env.py` re-pins `transformers==5.3.0` at startup (NeMo downgrades
-  it at build time) — it must run before any ML import in `app.py`.
+- `create_env.py` orchestrates a 4-step install at app startup — it must
+  run before any ML import in `app.py`:
+  1. Pin `huggingface-hub>=1.2,<2.0` (compatible with both gradio 6.20
+     and `nemo-toolkit`).
+  2. `pip install --no-deps nemo-toolkit[tts]==2.4.0` (avoiding a downgrade
+     of hub by the resolver).
+  3. Force-reinstall `transformers==5.3.0` (Qwen3.5 backbone that the
+     Gepard checkpoint was trained on).
+  4. Cap `numpy<2.0` so the codec/NeMo stack stays on numpy 1.x.
+- `nemo-toolkit` is installed at runtime (not in `requirements.txt`) to
+  keep gradio 6.20's `huggingface-hub>=1.2` constraint resolvable at
+  build time — NeMo's `transformers<=4.52` would otherwise pull hub<1.0.

app.py

@@ -1,9 +1,11 @@
 """GEPARD — text-to-speech inference Space (ZeroGPU), gradio.Server backend.
 
 Startup order is load-bearing:
-  1. ``create_env.setup_dependencies()`` re-pins transformers BEFORE any
-     ML import — the Space image ships whatever NeMo pinned at build time,
-     but the Gepard checkpoint requires transformers 5.3.0.
+  1. ``create_env.setup_dependencies()`` installs nemo-toolkit and re-pins
+     transformers BEFORE any ML import — the Space image ships gradio 6.20
+     which pins huggingface-hub>=1.2, and NeMo's transformers<=4.52 would
+     pull hub<1.0 at build time. create_env.py upgrades hub first, then
+     installs nemo with --no-deps, then re-pins transformers==5.3.0.
   2. ``spaces`` is imported before torch so ZeroGPU can patch CUDA calls.
   3. The engine (model + codec + speakers) is built ONCE at module level;
      ZeroGPU replays the recorded ``.to("cuda")`` moves when a GPU is

create_env.py

@@ -3,24 +3,62 @@ import subprocess
 import sys
 
 def setup_dependencies():
+    """Install NeMo and pin transformers in the right order to satisfy
+    gradio 6.20's huggingface-hub>=1.2 requirement while still ending up
+    on transformers 5.3.0 (which the Gepard checkpoint needs).
+
+    Why not put nemo-toolkit in requirements.txt? NeMo's
+    ``nemo-toolkit[tts]==2.4.0`` pulls in ``transformers<=4.52`` (which
+    resolves to 4.51.x), and that requires ``huggingface-hub<1.0`` —
+    conflicting with gradio 6.20's ``huggingface-hub>=1.2`` at build time
+    so pip refuses to resolve. We install NeMo at runtime instead, after
+    pinning hub and transformers to the versions we want.
+
+    Order is load-bearing:
+      1. transformers==5.3.0 with --no-deps — forces the runtime onto the
+         Qwen3.5 backbone version the Gepard checkpoint was trained on.
+         Pulls in hub>=1.2 transitively, satisfying gradio 6.20.
+      2. nemo-toolkit[tts]==2.4.0 with --no-deps — installed on top of
+         transformers 5.3.0 without re-resolving its declared (conflicting)
+         transformers<=4.52 constraint. The codec only needs
+         nemo.collections.tts.models.AudioCodecModel; the missing
+         transformers<5.x import paths NeMo *might* check for aren't
+         exercised at inference.
+      3. numpy<2.0 — keep the NeMo/codec stack on numpy 1.x.
+
+    Idempotent: writes a sentinel file so the install only runs once per
+    container.
+    """
     os.environ["OMP_NUM_THREADS"] = "4"
+
+    sentinel = "/tmp/deps_installed"
+    if os.path.exists(sentinel):
+        return
+
+    pip = [sys.executable, "-m", "pip", "install", "--no-cache-dir"]
     try:
-        if os.path.exists('/tmp/deps_installed'):
-            return
-            
-        # Re-pin transformers on top of whatever NeMo downgraded at build time
-        # (the Gepard checkpoint needs the Qwen3.5 backbone from 5.x). This
-        # upgrades huggingface-hub to 1.x — gradio 5.49 accepts <2.0, so the
-        # UI keeps working. numpy is capped so the force-reinstall does not
-        # bump it to 2.x under the NeMo stack built against 1.26.
-        print("Re-pinning transformers==5.3.0 ...")
-        subprocess.check_call([
-            sys.executable, "-m", "pip", "install", "--force-reinstall", "--no-cache-dir",
-            "transformers==5.3.0", "numpy<2.0"
+        # 1. Pin the transformers + hub versions gradio 6.20 needs, before
+        #    NeMo gets a chance to downgrade them. --no-deps because we
+        #    already have these libraries installed at compatible versions.
+        print("Pinning transformers==5.3.0 + hub>=1.2 (no-deps) ...")
+        subprocess.check_call(pip + [
+            "--no-deps", "transformers==5.3.0", "huggingface-hub>=1.2,<2.0",
         ])
-        
-        with open('/tmp/deps_installed', 'w') as f:
-            f.write('done')
-            
+
+        # 2. Install NeMo on top without re-resolving dependencies. Its
+        #    declared transformers<=4.52 / hub<1.0 would conflict; with
+        #    --no-deps the resolver doesn't see those constraints.
+        print("Installing nemo-toolkit[tts]==2.4.0 (no-deps) ...")
+        subprocess.check_call(pip + ["--no-deps", "nemo-toolkit[tts]==2.4.0"])
+
+        # 3. Cap numpy so a transitive bump doesn't break the NeMo/codec
+        #    stack, which is built against numpy 1.x.
+        print("Capping numpy<2.0 ...")
+        subprocess.check_call(pip + ["numpy<2.0"])
+
+        with open(sentinel, "w") as f:
+            f.write("done")
+
     except Exception as e:
-        print(f"Dependencies setup error: {e}")
+        print(f"Dependencies setup error: {e}")
+        raise

requirements.txt

@@ -1,12 +1,32 @@
 # GEPARD Space dependencies.
 #
 # torch is provided by the Space image — do NOT pin it here.
-nemo-toolkit[tts]==2.4.0
+# gradio is pinned via sdk_version in README.md (the Dockerfile installs it
+# with the [oauth,mcp] extras). Don't re-pin it here.
+#
+# nemo-toolkit[tts]==2.4.0 is installed in create_env.py at runtime instead
+# of here, because its `transformers<=4.52` constraint pulls in transformers
+# 4.51.x which requires huggingface-hub<1.0 — conflicting with gradio 6.20's
+# hub>=1.2 requirement at build time. create_env.py pins transformers==5.3.0
+# + hub>=1.2 first (with --no-deps), then installs nemo with --no-deps.
+# The deps listed below cover NeMo's runtime imports (nemo.collections.tts
+# only needs torch + omegaconf at import time, plus a few small libs).
 safetensors
 librosa>=0.10.0
 soundfile
-omegaconf
 pyyaml
-numpy
+numpy<2.0
 scipy
-gradio==6.20.0
+# NeMo runtime deps (not installed by `pip install nemo-toolkit[tts]` since
+# we use --no-deps to keep gradio 6.20's hub constraint resolvable).
+omegaconf
+webdataset
+pandas
+tqdm
+wget
+inflect
+num2words
+sentencepiece
+editdistance
+hydra-core
+packaging