Use bundled ZeroGPU Mamba wheel

Use the Hugging Face-provided mamba-ssm 2.3.2.post1 Blackwell wheel after validating the same package version locally with both Audex backends.

README.md

@@ -22,6 +22,8 @@ The demo includes **audio understanding, speech recognition, speech translation,
 
 Reasoning has no separate token cap by default for either model. `Max new tokens` sets a total cap shared by reasoning and the final answer.
 
+The hosted ZeroGPU runtime uses a prebuilt `mamba-ssm==2.3.2.post1` Blackwell wheel.
+
 ## Run locally
 
 ```bash

app.py

@@ -55,7 +55,7 @@ WHEELS_DIR = Path(os.environ.get("AUDEX_WHEELS_DIR", str(DEFAULT_WHEELS_DIR)))
 BUILD_DIR = Path(os.environ.get("AUDEX_BUILD_DIR", str(DEFAULT_BUILD_DIR)))
 REPO_ID = os.environ.get("SPACE_ID", "L0SG/Nemotron-Labs-Audex")
 CAUSAL_CONV1D_VERSION = "1.6.2.post1"
-MAMBA_SSM_VERSION = "2.3.1"
+MAMBA_SSM_VERSION = "2.3.2.post1"
 CAUSAL_CONV1D_SPEC = f"causal-conv1d=={CAUSAL_CONV1D_VERSION}"
 MAMBA_SSM_SPEC = f"mamba-ssm=={MAMBA_SSM_VERSION}"
 TORCH_ARCH = os.environ.get("AUDEX_TORCH_ARCH", "12.0" if IS_ZERO_GPU else "")