Add requirements.txt with pinned Jetson dependencies

README.md

@@ -93,10 +93,10 @@ Tested on Fleurs en_us samples — near-perfect output matching the fp16 baselin
 
 ### Option 1: Self-contained server (recommended for Jetson)
 
-No HuggingFace or vLLM dependencies needed. Runs inside the [PyTorch Jetson container](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/l4t-pytorch).
+No HuggingFace or vLLM dependencies needed. Tested on JetPack 6.x (R36.5.0), Python 3.10, CUDA 12.6.
 
 ```bash
-pip install safetensors websockets soundfile numpy librosa marlin
+pip install -r requirements.txt
 
 # Test with an audio file
 python scripts/jetson_serve_sdpa.py --test audio.wav

requirements.txt

@@ -0,0 +1,27 @@
+# Voxtral Mini 4B INT4 — Jetson Orin Nano (JetPack 6.x, R36.5.0)
+# Tested on: Python 3.10.12, CUDA 12.6, SM87 (Orin)
+#
+# Install PyTorch for Jetson first:
+#   pip install https://pypi.jetson-ai-lab.io/jp6/cu126/+f/02f/de421eabbf626/torch-2.9.1-cp310-cp310-linux_aarch64.whl
+#
+# Then install the rest:
+#   pip install -r requirements.txt
+
+# Core inference
+torch @ https://pypi.jetson-ai-lab.io/jp6/cu126/+f/02f/de421eabbf626/torch-2.9.1-cp310-cp310-linux_aarch64.whl
+safetensors==0.7.0
+numpy==2.2.6
+marlin @ git+https://github.com/IST-DASLab/marlin.git@1f25790bdd49fba53106164a24666dade68d7c90
+
+# Audio processing
+librosa==0.11.0
+soundfile==0.13.1
+
+# Tokenizer
+mistral_common==1.9.1
+
+# WebSocket server
+websockets>=12.0
+
+# CUDA runtime (pulled in by torch, listed for reference)
+nvidia-cudss-cu12==0.7.1.6