Repository navigation
Refactor exporting NeMo models - #2362
Conversation
WalkthroughThe changes introduce support for int8 quantized ONNX models across multiple NeMo ASR model export workflows, scripts, and documentation. New workflow steps automate quantization, packaging, publishing, and testing of these int8 models. Additional model variants and support for the Parakeet TDT CTC model (Japanese) are also added, with corresponding scripts and integration into the Kotlin API. Changes
Sequence Diagram(s)sequenceDiagram
participant Workflow
participant ExportScript
participant ONNX
participant Quantizer
participant Archive
participant HuggingFace
Workflow->>ExportScript: Trigger export (model, tokens)
ExportScript->>ONNX: Export model to model.onnx
ExportScript->>Quantizer: Quantize model.onnx to model.int8.onnx
ExportScript->>Archive: Package models and tokens
Workflow->>HuggingFace: Publish original and int8 models
sequenceDiagram
participant User
participant RunScript
participant ModelDir
participant Tester
User->>RunScript: Start run-ctc.sh / run-transducer.sh
RunScript->>ModelDir: Organize original and int8 models
RunScript->>Tester: Test original model
RunScript->>Tester: Test int8 model
Poem
📜 Recent review detailsConfiguration used: CodeRabbit UI 📒 Files selected for processing (1)
🚧 Files skipped from review as they are similar to previous changes (1)
✨ Finishing Touches
🪧 TipsChatThere are 3 ways to chat with CodeRabbit:
SupportNeed help? Create a ticket on our support page for assistance with any issues or questions. Note: Be mindful of the bot's finite context window. It's strongly recommended to break down tasks such as reading entire modules into smaller chunks. For a focused discussion, use review comments to chat about specific files and their changes, instead of using the PR comments. CodeRabbit Commands (Invoked using PR comments)
Other keywords and placeholders
CodeRabbit Configuration File (
|
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 4
♻️ Duplicate comments (2)
.github/workflows/export-nemo-fast-conformer-hybrid-transducer-ctc.yaml (2)
65-68: Same quoting issue as transducer workflowPlease apply the
${names[@]}→"${names[@]}"quoting fix here as well.
101-116: Replicate the idempotency / commit-message improvements hereThe comments made for the transducer workflow’s publish step equally apply to this CTC workflow.
🧹 Nitpick comments (11)
scripts/nemo/fast-conformer-hybrid-transducer-ctc/export-onnx-ctc-non-streaming.py (1)
90-94: Consider adding error handling for quantization step.The dynamic quantization logic is correctly implemented and placed after metadata addition. However, consider adding error handling around the quantization step to provide better feedback if the quantization fails.
+ try: quantize_dynamic( model_input="./model.onnx", model_output="./model.int8.onnx", weight_type=QuantType.QUInt8, ) + print("Quantized model saved to model.int8.onnx") + except Exception as e: + print(f"Warning: Quantization failed: {e}") + print("Continuing with original model only")scripts/nemo/fast-conformer-hybrid-transducer-ctc/export-onnx-transducer-non-streaming.py (1)
94-99: Consider adding error handling and improving variable naming.The dynamic quantization logic is correctly implemented for all three model components. However, consider adding error handling and using a more descriptive variable name.
- for m in ["encoder", "decoder", "joiner"]: + for model_name in ["encoder", "decoder", "joiner"]: + try: quantize_dynamic( - model_input=f"{m}.onnx", - model_output=f"{m}.int8.onnx", + model_input=f"{model_name}.onnx", + model_output=f"{model_name}.int8.onnx", weight_type=QuantType.QUInt8, ) + print(f"Quantized {model_name} model saved to {model_name}.int8.onnx") + except Exception as e: + print(f"Warning: Quantization failed for {model_name}: {e}")scripts/nemo/fast-conformer-hybrid-transducer-ctc/export-onnx-transducer.py (1)
126-131: Consider adding error handling and improving variable naming.The dynamic quantization logic is correctly implemented for all three model components. However, consider adding error handling and using a more descriptive variable name.
- for m in ["encoder", "decoder", "joiner"]: + for model_name in ["encoder", "decoder", "joiner"]: + try: quantize_dynamic( - model_input=f"{m}.onnx", - model_output=f"{m}.int8.onnx", + model_input=f"{model_name}.onnx", + model_output=f"{model_name}.int8.onnx", weight_type=QuantType.QUInt8, ) + print(f"Quantized {model_name} model saved to {model_name}.int8.onnx") + except Exception as e: + print(f"Warning: Quantization failed for {model_name}: {e}")scripts/nemo/fast-conformer-hybrid-transducer-ctc/run-transducer.sh (1)
26-31: Consider adding error handling for file operations.The file organization logic is correct, but consider adding error handling to ensure robustness when moving files.
cp -v tokens.txt $d/ + # Check if int8 models exist before moving + if [ -f encoder.int8.onnx ] && [ -f decoder.int8.onnx ] && [ -f joiner.int8.onnx ]; then mv -v encoder.int8.onnx $d_int8/ mv -v decoder.int8.onnx $d_int8/ mv -v joiner.int8.onnx $d_int8/ mv -v tokens.txt $d_int8/ + else + echo "Warning: Some int8 models not found, skipping int8 directory setup" + fiscripts/nemo/fast-conformer-hybrid-transducer-ctc/run-ctc.sh (2)
17-19: Quote array expansions to silence ShellCheck SC2068ShellCheck flags the un-quoted array expansions:
for m in ${ms[@]}; do # SC2068 ... for m in ${ms[@]}; do # SC2068Use double-quotes to avoid word-splitting and future-proof the loop:
-for m in ${ms[@]}; do +for m in "${ms[@]}"; doAlso applies to: 41-42
24-30: Streamlinetokens.txthandling
tokens.txtis first copied to the FP dir (cp) and then moved to the int8 dir (mv).
A simpler, clearer flow is to copy into both targets and leave the working copy untouched:- cp -v tokens.txt $d/ - ... - mv -v tokens.txt $d_int8/ + cp -v tokens.txt $d/ + cp -v tokens.txt $d_int8/This avoids mixing
cp/mv, keeps a local copy for diagnostics, and avoids re-creating the file on every loop iteration.scripts/nemo/parakeet-tdt_ctc-0.6b-ja/run-ctc.sh (1)
1-4: Adoptset -euo pipefailfor safer bash scriptsReplacing
set -exwith the stricter variant catches unset variables (-u) and pipe errors (-o pipefail):-set -ex +set -euo pipefail +set -x # keep trace if desiredThis small hardening step prevents subtle failures from passing silently.
.github/workflows/export-nemo-fast-conformer-hybrid-transducer-transducer.yaml (1)
65-68: Quote the${models[@]}expansion inside the loopSame SC2068 issue as in the shell scripts – quoting avoids accidental word-splitting if a future model name contains spaces:
-for m in ${models[@]}; do +for m in "${models[@]}"; doscripts/nemo/fast-conformer-hybrid-transducer-ctc/run-ctc-non-streaming.sh (1)
171-187: Quote list variables inside thefor w in …loopsAlthough the file names are plain ASCII, quoting is a good habit and prevents edge-cases:
-for w in en-english.wav de-german.wav ...; do +for w in en-english.wav de-german.wav ...; do # <<< quote "$w" below ... - --wav $data/$w + --wav "$data/$w"This pattern repeats for the int8 loop; fix both.
scripts/nemo/parakeet-tdt_ctc-0.6b-ja/export-onnx-ctc.py (2)
74-74: Replace os.system with subprocess for better error handling.Using
os.systemfor file operations doesn't provide good error handling and is generally discouraged.- os.system("ls -lh *.onnx") + import subprocess + subprocess.run(["ls", "-lh", "*.onnx"], check=True)Apply the same change to line 84.
Also applies to: 84-84
35-37: Consider parameterizing the hardcoded model name.The model name is hardcoded, which limits reusability. Consider making it configurable through command-line arguments.
+import argparse + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument("--model", default="nvidia/parakeet-tdt_ctc-0.6b-ja") + return parser.parse_args() + @torch.no_grad() def main(): + args = get_args() asr_model = nemo_asr.models.ASRModel.from_pretrained( - model_name="nvidia/parakeet-tdt_ctc-0.6b-ja" + model_name=args.model )
📜 Review details
Configuration used: CodeRabbit UI
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (19)
.github/workflows/export-nemo-fast-conformer-hybrid-transducer-ctc-non-streaming.yaml(2 hunks).github/workflows/export-nemo-fast-conformer-hybrid-transducer-ctc.yaml(2 hunks).github/workflows/export-nemo-fast-conformer-hybrid-transducer-transducer-non-streaming.yaml(2 hunks).github/workflows/export-nemo-fast-conformer-hybrid-transducer-transducer.yaml(2 hunks).github/workflows/export-nemo-parakeet-tdt.yaml(1 hunks)scripts/apk/generate-vad-asr-apk-script.py(1 hunks)scripts/nemo/fast-conformer-hybrid-transducer-ctc/README.md(1 hunks)scripts/nemo/fast-conformer-hybrid-transducer-ctc/export-onnx-ctc-non-streaming.py(2 hunks)scripts/nemo/fast-conformer-hybrid-transducer-ctc/export-onnx-ctc.py(2 hunks)scripts/nemo/fast-conformer-hybrid-transducer-ctc/export-onnx-transducer-non-streaming.py(2 hunks)scripts/nemo/fast-conformer-hybrid-transducer-ctc/export-onnx-transducer.py(2 hunks)scripts/nemo/fast-conformer-hybrid-transducer-ctc/run-ctc-non-streaming.sh(10 hunks)scripts/nemo/fast-conformer-hybrid-transducer-ctc/run-ctc.sh(1 hunks)scripts/nemo/fast-conformer-hybrid-transducer-ctc/run-transducer-non-streaming.sh(10 hunks)scripts/nemo/fast-conformer-hybrid-transducer-ctc/run-transducer.sh(2 hunks)scripts/nemo/parakeet-tdt_ctc-0.6b-ja/export-onnx-ctc.py(1 hunks)scripts/nemo/parakeet-tdt_ctc-0.6b-ja/run-ctc.sh(1 hunks)scripts/nemo/parakeet-tdt_ctc-0.6b-ja/test-onnx-ctc-non-streaming.py(1 hunks)sherpa-onnx/kotlin-api/OfflineRecognizer.kt(1 hunks)
🧰 Additional context used
🧬 Code Graph Analysis (1)
scripts/nemo/parakeet-tdt_ctc-0.6b-ja/export-onnx-ctc.py (2)
scripts/nemo/parakeet-tdt_ctc-0.6b-ja/test-onnx-ctc-non-streaming.py (1)
main(107-165)scripts/spleeter/export_onnx.py (1)
export(47-74)
🪛 LanguageTool
scripts/nemo/fast-conformer-hybrid-transducer-ctc/README.md
[grammar] ~26-~26: Use correct spacing
Context: ...s/nemo/models/parakeet-tdt_ctc-110m - https://huggingface.co/nvidia/parakeet-tdt_ctc-0.6b-ja to sherpa-onnx.
(QB_NEW_EN_OTHER_ERROR_IDS_5)
🪛 markdownlint-cli2 (0.17.2)
scripts/nemo/fast-conformer-hybrid-transducer-ctc/README.md
26-26: Unordered list indentation
Expected: 0; Actual: 2
(MD007, ul-indent)
26-26: Bare URL used
(MD034, no-bare-urls)
🪛 Shellcheck (0.10.0)
scripts/nemo/fast-conformer-hybrid-transducer-ctc/run-ctc.sh
[error] 41-41: Double quote array expansions to avoid re-splitting elements.
(SC2068)
🪛 actionlint (1.7.7)
.github/workflows/export-nemo-parakeet-tdt.yaml
54-54: shellcheck reported issue in this script: SC2068:error:4:10: Double quote array expansions to avoid re-splitting elements
(shellcheck)
54-54: shellcheck reported issue in this script: SC2086:info:5:47: Double quote to prevent globbing and word splitting
(shellcheck)
54-54: shellcheck reported issue in this script: SC2086:info:6:12: Double quote to prevent globbing and word splitting
(shellcheck)
54-54: shellcheck reported issue in this script: SC2086:info:6:23: Double quote to prevent globbing and word splitting
(shellcheck)
🔇 Additional comments (21)
scripts/nemo/parakeet-tdt_ctc-0.6b-ja/test-onnx-ctc-non-streaming.py (2)
2-169: LGTM! Well-structured test implementation.The Python implementation is well-organized with proper error handling, feature extraction, and CTC decoding logic. The code follows good practices for ONNX model testing.
1-1: Remove the confusing script reference.Line 1 contains a reference to another script file, but the rest of the file contains a complete Python implementation. This creates confusion about the file's purpose. If this is meant to be a standalone test script, remove the reference line.
Apply this diff to remove the problematic reference:
-../fast-conformer-hybrid-transducer-ctc/test-onnx-ctc-non-streaming.pyLikely an incorrect or invalid review comment.
.github/workflows/export-nemo-fast-conformer-hybrid-transducer-transducer-non-streaming.yaml (2)
64-68: LGTM! Consistent int8 model variant addition.The int8 model variants are properly added to the publishing pipeline with clear naming conventions.
96-100: LGTM! Matching compression configuration.The int8 model variants are consistently added to the compression step, maintaining alignment with the publishing configuration.
scripts/nemo/fast-conformer-hybrid-transducer-ctc/export-onnx-ctc.py (2)
9-9: LGTM! Proper quantization imports.The necessary imports for dynamic quantization are correctly added.
118-122: LGTM! Well-placed quantization step.The dynamic quantization is properly implemented with appropriate weight type and positioned correctly in the workflow after metadata addition.
.github/workflows/export-nemo-fast-conformer-hybrid-transducer-ctc-non-streaming.yaml (2)
64-68: LGTM! Consistent int8 model integration.The int8 model variants are properly integrated into the publishing workflow with consistent naming patterns.
97-101: LGTM! Aligned compression configuration.The int8 model variants are consistently added to the compression step, maintaining proper alignment with the publishing workflow.
scripts/nemo/fast-conformer-hybrid-transducer-ctc/export-onnx-ctc-non-streaming.py (1)
9-9: LGTM - Quantization import added correctly.The import of
QuantTypeandquantize_dynamicfromonnxruntime.quantizationis properly added to support dynamic quantization functionality.scripts/nemo/fast-conformer-hybrid-transducer-ctc/export-onnx-transducer-non-streaming.py (1)
9-9: LGTM - Quantization import added correctly.The import of
QuantTypeandquantize_dynamicfromonnxruntime.quantizationis properly added to support dynamic quantization functionality.scripts/nemo/fast-conformer-hybrid-transducer-ctc/export-onnx-transducer.py (1)
9-9: LGTM - Quantization import added correctly.The import of
QuantTypeandquantize_dynamicfromonnxruntime.quantizationis properly added to support dynamic quantization functionality.scripts/nemo/fast-conformer-hybrid-transducer-ctc/run-transducer.sh (2)
20-22: LGTM - Directory structure for int8 models.The addition of separate directories for original and int8 quantized models is well-organized and maintains clear separation between model variants.
52-58: LGTM - Testing both model variants.The testing logic correctly validates both original and int8 quantized models, ensuring parity between the two variants.
sherpa-onnx/kotlin-api/OfflineRecognizer.kt (2)
605-613: LGTM - English Parakeet TDT CTC model configuration.The new model configuration for the English Parakeet TDT CTC model is properly structured and follows the established pattern for NeMo CTC models.
615-623: LGTM - Japanese Parakeet TDT CTC model configuration.The new model configuration for the Japanese Parakeet TDT CTC model is properly structured and follows the established pattern for NeMo CTC models.
scripts/nemo/fast-conformer-hybrid-transducer-ctc/run-ctc-non-streaming.sh (1)
21-29: Duplicatemv tokens.txtlikely unintended
tokens.txtis first copied to the FP directory (line 22) and then moved there again at line 28 – the secondmvis redundant and can fail if the file was already moved elsewhere earlier in the script.-cp -v tokens.txt $d/ -... -mv -v tokens.txt $d/Remove the second line (or switch both to
cp) for clarity.Likely an incorrect or invalid review comment.
scripts/nemo/fast-conformer-hybrid-transducer-ctc/run-transducer-non-streaming.sh (2)
21-33: Well-structured int8 model organization.The systematic approach of creating separate directories for regular and int8 models with explicit file moves ensures proper organization and avoids potential conflicts.
154-172: Comprehensive testing coverage for int8 models.The addition of int8 model testing maintains parity with the regular model testing, ensuring both variants are properly validated.
scripts/nemo/parakeet-tdt_ctc-0.6b-ja/export-onnx-ctc.py (1)
12-30: Metadata handling looks correct.The function properly clears existing metadata before adding new entries, ensuring clean metadata state.
.github/workflows/export-nemo-parakeet-tdt.yaml (2)
64-95: Well-implemented publishing workflow with retry mechanism.The Hugging Face publishing step includes proper authentication, retry logic, and Git LFS handling for large files.
97-105: Proper release asset handling.The release step correctly uses file globbing and proper authentication for uploading to the sherpa-onnx repository.
| - https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/stt_multilingual_fastconformer_hybrid_large_pc | ||
|
|
||
| - https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/parakeet-tdt_ctc-110m | ||
| - https://huggingface.co/nvidia/parakeet-tdt_ctc-0.6b-ja |
There was a problem hiding this comment.
Fix markdown formatting issues.
The line has incorrect indentation and should format the URL properly according to markdown standards.
Apply this diff to fix the formatting:
- - https://huggingface.co/nvidia/parakeet-tdt_ctc-0.6b-ja
+- https://huggingface.co/nvidia/parakeet-tdt_ctc-0.6b-jaCommittable suggestion skipped: line range outside the PR's diff.
🧰 Tools
🪛 LanguageTool
[grammar] ~26-~26: Use correct spacing
Context: ...s/nemo/models/parakeet-tdt_ctc-110m - https://huggingface.co/nvidia/parakeet-tdt_ctc-0.6b-ja to sherpa-onnx.
(QB_NEW_EN_OTHER_ERROR_IDS_5)
🪛 markdownlint-cli2 (0.17.2)
26-26: Unordered list indentation
Expected: 0; Actual: 2
(MD007, ul-indent)
26-26: Bare URL used
(MD034, no-bare-urls)
🤖 Prompt for AI Agents
In scripts/nemo/fast-conformer-hybrid-transducer-ctc/README.md at line 26, the
URL is not properly formatted in markdown due to incorrect indentation. Fix this
by removing any leading spaces before the URL and ensure it is either on its own
line or formatted as a proper markdown link so it renders correctly.
| for m in ${models[@]}; do | ||
| rm -rf huggingface | ||
| export GIT_LFS_SKIP_SMUDGE=1 | ||
| export GIT_CLONE_PROTECTION_ACTIVE=false | ||
| git clone https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/$m huggingface | ||
| cp -av $m/* huggingface | ||
| cd huggingface | ||
| git lfs track "*.onnx" | ||
| git lfs track "*.wav" | ||
| git status | ||
| git add . | ||
| git status | ||
| git commit -m "first commit" | ||
| git push https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/$m main | ||
| cd .. | ||
| done |
There was a problem hiding this comment.
🛠️ Refactor suggestion
Make the Hugging Face publish loop idempotent & cheaper
git lfs track "*.onnx"and"*.wav"run on every iteration – do it once outside the loop or inside the repo after the first clone.- The commit message
"first commit"becomes inaccurate after the first push – consider a generic message ("update models") or includedate +"%F"to avoid duplicate-headline warnings. - Add
set -euo pipefailat the top of the script block to fail fast on any publishing error.
No diff provided since it spans multiple script blocks.
🤖 Prompt for AI Agents
In
.github/workflows/export-nemo-fast-conformer-hybrid-transducer-transducer.yaml
around lines 101 to 116, the git lfs track commands for "*.onnx" and "*.wav" are
redundantly run inside the loop for each model; move these commands outside the
loop or run them only once after the initial clone to improve efficiency. Also,
update the commit message from the static "first commit" to a more generic
message like "update models" or include a timestamp using date +"%F" to prevent
duplicate commit message warnings. Finally, add set -euo pipefail at the top of
the script block to ensure the script fails immediately on any error during
publishing.
| run: | | ||
| models=( | ||
| sherpa-onnx-nemo-parakeet-tdt_ctc-0.6b-ja-35000-int8 | ||
| ) | ||
| for m in ${models[@]}; do | ||
| mv -v scripts/nemo/parakeet-tdt_ctc-0.6b-ja/$m . | ||
| tar cjfv $m.tar.bz2 $m | ||
| done |
There was a problem hiding this comment.
🛠️ Refactor suggestion
Fix shell script quoting issues.
Shellcheck identifies several quoting issues that should be addressed for robustness.
models=(
sherpa-onnx-nemo-parakeet-tdt_ctc-0.6b-ja-35000-int8
)
- for m in ${models[@]}; do
- mv -v scripts/nemo/parakeet-tdt_ctc-0.6b-ja/$m .
- tar cjfv $m.tar.bz2 $m
+ for m in "${models[@]}"; do
+ mv -v "scripts/nemo/parakeet-tdt_ctc-0.6b-ja/$m" .
+ tar cjfv "$m.tar.bz2" "$m"
done📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| run: | | |
| models=( | |
| sherpa-onnx-nemo-parakeet-tdt_ctc-0.6b-ja-35000-int8 | |
| ) | |
| for m in ${models[@]}; do | |
| mv -v scripts/nemo/parakeet-tdt_ctc-0.6b-ja/$m . | |
| tar cjfv $m.tar.bz2 $m | |
| done | |
| run: | | |
| models=( | |
| sherpa-onnx-nemo-parakeet-tdt_ctc-0.6b-ja-35000-int8 | |
| ) | |
| for m in "${models[@]}"; do | |
| mv -v "scripts/nemo/parakeet-tdt_ctc-0.6b-ja/$m" . | |
| tar cjfv "$m.tar.bz2" "$m" | |
| done |
🧰 Tools
🪛 actionlint (1.7.7)
54-54: shellcheck reported issue in this script: SC2068:error:4:10: Double quote array expansions to avoid re-splitting elements
(shellcheck)
54-54: shellcheck reported issue in this script: SC2086:info:5:47: Double quote to prevent globbing and word splitting
(shellcheck)
54-54: shellcheck reported issue in this script: SC2086:info:6:12: Double quote to prevent globbing and word splitting
(shellcheck)
54-54: shellcheck reported issue in this script: SC2086:info:6:23: Double quote to prevent globbing and word splitting
(shellcheck)
🤖 Prompt for AI Agents
In .github/workflows/export-nemo-parakeet-tdt-tdt.yaml around lines 54 to 61,
the shell script has quoting issues that can cause word splitting or globbing
problems. Fix this by quoting the array expansion as "${models[@]}" in the for
loop and quoting variable expansions like "$m" in the mv and tar commands to
ensure proper handling of filenames and prevent errors.
There was a problem hiding this comment.
Pull Request Overview
Refactors and extends model export support to include new NeMo Parakeet TDT int8 variants for English and Japanese, updating the Kotlin API, export scripts, test runners, and CI workflows.
- Added support for two new int8 model types in
OfflineRecognizer.kt. - Enhanced Python export scripts to perform dynamic quantization and metadata injection.
- Updated shell scripts and GitHub workflows to package, test, and publish int8 model artifacts.
Reviewed Changes
Copilot reviewed 19 out of 19 changed files in this pull request and generated 1 comment.
Show a summary per file
| File | Description |
|---|---|
| sherpa-onnx/kotlin-api/OfflineRecognizer.kt | Added cases 33/34 for new int8 Parakeet TDT ASR models |
| scripts/nemo/parakeet-tdt_ctc-0.6b-ja/test-onnx-ctc-non-streaming.py | Introduced a placeholder script referencing the test runner |
| scripts/nemo/parakeet-tdt_ctc-0.6b-ja/export-onnx-ctc.py | Imported quantize_dynamic, added int8 quantization step |
| scripts/nemo/fast-conformer-hybrid-transducer-ctc/*.sh | Extended CTC/transducer run scripts to handle int8 models |
| scripts/nemo/fast-conformer-hybrid-transducer-ctc/*.py | Imported quantization, looped over encoder/decoder/joiner |
| .github/workflows/*.yaml | Updated workflows to package and publish int8 variants |
Comments suppressed due to low confidence (2)
sherpa-onnx/kotlin-api/OfflineRecognizer.kt:616
- The directory name for the Japanese int8 model uses a different separator pattern (
parakeet-tdt_ctc) compared to the English int8 model (parakeet_tdt_ctc), which may lead to inconsistency or path resolution errors. Consider aligning the naming convention acrossmodelDirvalues.
val modelDir = "sherpa-onnx-nemo-parakeet-tdt_ctc-0.6b-ja-35000-int8"
sherpa-onnx/kotlin-api/OfflineRecognizer.kt:605
- [nitpick] New model type indices (33 and 34) have been added, but there is no corresponding unit test to verify that
getOfflineModelConfigreturns the expectedOfflineModelConfigfor these indices. Consider adding tests to cover these new cases.
33 -> {
See also
Summary by CodeRabbit
New Features
Documentation
Chores