fix deps

[yqwangustc]

To evaluate audio/speech, we need to install ffmpeg and torchcodec

Summary by CodeRabbit

• Chores
  • Added FFmpeg and torchcodec audio processing dependencies to the project configuration.

[coderabbitai]

📝 Walkthrough

Walkthrough

This PR adds FFmpeg support to the NeMo Skills project by installing the ffmpeg system package in the Docker container and adding the torchcodec Python dependency with FFmpeg configuration documentation.

Changes

Cohort / File(s)	Summary
FFmpeg/Audio Support Infrastructure dockerfiles/Dockerfile.nemo-skills, requirements/audio.txt	Adds ffmpeg system package installation in Docker and introduces torchcodec Python dependency with FFmpeg setup notes to enable audio processing capabilities.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~5 minutes

🚥 Pre-merge checks | ✅ 2 | ❌ 1

❌ Failed checks (1 inconclusive)

Check name	Status	Explanation	Resolution
Title check	❓ Inconclusive	The title 'fix deps' is vague and generic, using non-descriptive language that doesn't clearly communicate the specific changes being made.	Consider using a more descriptive title like 'Add ffmpeg and torchcodec dependencies for audio/speech evaluation' to clearly indicate what dependencies are being added.

✅ Passed checks (2 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment

Tip

Issue Planner is now in beta. Read the docs and try it out! Share your feedback on Discord.

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 1

🧹 Nitpick comments (1)

requirements/audio.txt (1)

9-9: torchcodec is unpinned; each release is compatible with only a specific torch version.

For other versions, refer to the table for compatibility between versions of torch and torchcodec. Without a version constraint, pip install will pull the latest release on every fresh install, which could resolve to a torchcodec version that is incompatible with the torch version already in the environment. All other entries in this file carry at least a lower-bound constraint (e.g., jiwer>=3.1.0,<4.0.0); torchcodec should follow the same pattern.

♻️ Suggested change

-torchcodec
+torchcodec>=0.5.0  # pin lower-bound; check torch<->torchcodec compatibility table before bumping

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@requirements/audio.txt` at line 9, The requirements entry for the package
identifier "torchcodec" is unpinned and risks installing a release incompatible
with the project's installed "torch" version; update the requirements entry for
torchcodec to include a version constraint that matches your torch release (use
the torch/torchcodec compatibility table to pick the correct semver), e.g., add
a lower-bound and upper-bound (like >=X.Y.Z,<X.Y+1) or pin to an exact
compatible release; ensure you update the "torchcodec" line in
requirements/audio.txt (the package name "torchcodec") accordingly and comment
or document which torch version it targets.

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@requirements/audio.txt`:
- Around line 5-9: The requirements entry for torchcodec silently installs
CPU-only wheels; update the torchcodec line in requirements/audio.txt to include
a clear inline note warning that GPU users must install via the PyTorch wheel
index (e.g., use pip install torchcodec --index-url
https://download.pytorch.org/whl/<cuda_suffix>) and add a version pin to match
the project's PyTorch version (e.g., pin to the same minor/patch compatibility
as torch) and also add a short README or installation-guide entry explaining how
to choose the correct <cuda_suffix> to match torch; target the existing
"torchcodec" line so reviewers can find and update it.

---

Nitpick comments:
In `@requirements/audio.txt`:
- Line 9: The requirements entry for the package identifier "torchcodec" is
unpinned and risks installing a release incompatible with the project's
installed "torch" version; update the requirements entry for torchcodec to
include a version constraint that matches your torch release (use the
torch/torchcodec compatibility table to pick the correct semver), e.g., add a
lower-bound and upper-bound (like >=X.Y.Z,<X.Y+1) or pin to an exact compatible
release; ensure you update the "torchcodec" line in requirements/audio.txt (the
package name "torchcodec") accordingly and comment or document which torch
version it targets.

[coderabbitai]

⚠️ Potential issue | 🟠 Major

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
# Find all requirements files and check for torch version constraints
find . -name "*.txt" -path "*/requirements/*" -type f | head -20

Repository: NVIDIA-NeMo/Skills

Length of output: 254

---

🏁 Script executed:

#!/bin/bash
# Search for torch version constraints in requirements files
rg "torch" requirements/ --type txt -i

Repository: NVIDIA-NeMo/Skills

Length of output: 253

---

🏁 Script executed:

#!/bin/bash
# Check the structure of requirements directory
ls -la requirements/ 2>/dev/null || echo "No requirements directory at root"

Repository: NVIDIA-NeMo/Skills

Length of output: 755

---

🏁 Script executed:

#!/bin/bash
# Look for any documentation about installation/setup
find . -name "README*" -o -name "INSTALL*" -o -name "setup.py" -o -name "pyproject.toml" | head -20

Repository: NVIDIA-NeMo/Skills

Length of output: 504

---

🏁 Script executed:

#!/bin/bash
# Check the audio.txt file to see all dependencies and version constraints
cat -n requirements/audio.txt 2>/dev/null || echo "File not found"

Repository: NVIDIA-NeMo/Skills

Length of output: 610

---

🏁 Script executed:

#!/bin/bash
# Read main.txt to see torch version constraints
cat -n requirements/main.txt

Repository: NVIDIA-NeMo/Skills

Length of output: 1934

---

🏁 Script executed:

#!/bin/bash
# Check README for installation/setup instructions
cat README.md | head -100

Repository: NVIDIA-NeMo/Skills

Length of output: 8337

---

🏁 Script executed:

#!/bin/bash
# Look for Dockerfile or setup scripts that might handle GPU/CUDA installation
find . -name "Dockerfile*" -o -name "*setup*" -o -name "install*" | grep -v ".git" | head -20

Repository: NVIDIA-NeMo/Skills

Length of output: 341

---

🏁 Script executed:

#!/bin/bash
# Check if there's a Dockerfile that might document CUDA/GPU setup
find . -name "Dockerfile*" -type f

Repository: NVIDIA-NeMo/Skills

Length of output: 258

---

🌐 Web query:

torchcodec PyPI installation CPU GPU CUDA index-url

💡 Result:

• CPU-only (from PyPI):

  pip install torchcodec

  This installs the CPU-only wheels by default. [1]

• CUDA/GPU (Linux, from the PyTorch wheel index): install from the PyTorch CUDA-specific index matching the CUDA build you used for PyTorch, e.g. CUDA 12.6:

  pip install torchcodec --index-url https://download.pytorch.org/whl/cu126

  If you don’t pass --index-url, pip will install the CPU-only build. [2]

• CUDA/GPU (Windows): GPU support is via conda-forge (experimental/beta):

  conda install -c conda-forge "torchcodec=*=*cuda*"

  [2]

Notes: TorchCodec relies on having FFmpeg installed (and for CUDA decoding, FFmpeg must be built with NVDEC support). [1][2]

Sources: [1] PyPI torchcodec page, [2] torchcodec GitHub README (Installing CUDA-enabled TorchCodec).

---

pip install -r requirements/audio.txt silently installs the CPU-only build of torchcodec.

Without the --index-url flag, pip defaults to the CPU-only wheels from PyPI. For GPU-accelerated decoding, the correct installation requires:

pip install torchcodec --index-url https://download.pytorch.org/whl/cu126

The CUDA suffix must match your PyTorch installation. Since pip requirements files cannot embed custom index URLs, add a clear note on the torchcodec line to warn users, and consider documenting the full GPU installation procedure in the README or a dedicated installation guide. The current comments only describe the FFmpeg system dependency, not the PyPI index requirement.

Additionally, add a version constraint to torchcodec to pin it alongside torch. Like other dependencies in this file, torchcodec has tight version coupling with PyTorch, and an unpinned version could break on future releases.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@requirements/audio.txt` around lines 5 - 9, The requirements entry for
torchcodec silently installs CPU-only wheels; update the torchcodec line in
requirements/audio.txt to include a clear inline note warning that GPU users
must install via the PyTorch wheel index (e.g., use pip install torchcodec
--index-url https://download.pytorch.org/whl/<cuda_suffix>) and add a version
pin to match the project's PyTorch version (e.g., pin to the same minor/patch
compatibility as torch) and also add a short README or installation-guide entry
explaining how to choose the correct <cuda_suffix> to match torch; target the
existing "torchcodec" line so reviewers can find and update it.