eustlb

out/
__pycache__/
.media/

eustlb

out/
__pycache__/
.media/

eustlb

out/
__pycache__/
.media/

eustlb

"""Minimal demo: {% generation %} -> assistant_masks -> labels -> loss, Llama 3.1 8B.

Run with --forward to also load the model and compute the masked loss.
"""
import sys
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

MODEL = "meta-llama/Llama-3.1-8B-Instruct"

eustlb

diff --git a/tests/models/embedding_gemma2/test_modeling_embedding_gemma2.py b/tests/models/embedding_gemma2/test_modeling_embedding_gemma2.py
index 9fb1f65e87..d4ad39242b 100644
--- a/tests/models/embedding_gemma2/test_modeling_embedding_gemma2.py
+++ b/tests/models/embedding_gemma2/test_modeling_embedding_gemma2.py
@@ -464,18 +464,19 @@ class EmbeddingGemma2ModelTest(ModelTesterMixin, unittest.TestCase):
 class EmbeddingGemma2IntegrationTest(MemoryCleanupMixin, unittest.TestCase):
     """
     reproducer (uploads the golden to
-    ``hf://buckets/hf-internal-testing/embeddinggemma2-integration-test/<case>/expected_embeddings.safetensors``,
+    ``hf://buckets/hf-internal-testing/embeddinggemma2-integration-tests/<case>/expected_embeddings.safetensors``,

eustlb

"""Regenerates the fixtures used by `GraniteSpeech5ForCTCIntegrationTest`.

    python tests/models/granite_speech5/reproducer.py

Everything comes from the original remote-code implementation, so the fixtures are an independent
reference rather than a snapshot of the port's own output.
"""

import json
import pathlib

eustlb

[project]
name = "granite-speech-nar-reproducers"
version = "0.0.0"
description = "Reference reproducers for the HF GraniteSpeechNarForCTC test fixtures, from the original remote-code model."

# Pinned to the transformers version the original checkpoint was authored against
# (`config.json` -> `transformers_version: 5.8.0.dev0`), so the remote-code model reproduces exactly.
# An isolated env is required: the transformers dev checkout under test natively registers the
# `granite_speech_nar` model_type, which would clash with the repo's `trust_remote_code` classes.
requires-python = ">=3.10,<3.13"

eustlb

"""Capture the HF Nemotron3_5AsrForRNNT outputs to bake into the integration tests.

Not a reproducer (those are the NeMo reference). This runs the *HF* model on the same audio the tests
use, conditioned on the language prompt `language="en-US"` (the test samples are English):

- offline single (librispeech_asr_dummy sample 0, default att_context [56, 3])
- offline batched (librispeech_asr_dummy samples 0..4, default att_context [56, 3])
- streaming (obama.mp3, num_lookahead_tokens=6 i.e. att_context [56, 6], mel-frame chunks 49 then 56)
"""

eustlb

"""Capture the HF NemotronAsrForRNNT outputs to bake into the integration tests.

Not a reproducer (those are NeMo reference). This runs the *HF* model on the same audio the tests use:
  - offline single   (librispeech_asr_dummy sample 0, default att_context)
  - offline batched   (librispeech_asr_dummy samples 0..4)
  - streaming         (obama.mp3, att_context_size=[70, 6], mel-frame chunks 49 then 56)
and prints JSON so the exact HF strings can be pasted into EXPECTED_* in the test.
"""

import json

eustlb

"""Capture the HF NemotronAsrForRNNT outputs to bake into the integration tests.

Not a reproducer (those are NeMo reference). This runs the *HF* model on the same audio the tests use:
  - offline single   (librispeech_asr_dummy sample 0, default att_context)
  - offline batched   (librispeech_asr_dummy samples 0..4)
  - streaming         (obama.mp3, att_context_size=[70, 6], mel-frame chunks 49 then 56)
and prints JSON so the exact HF strings can be pasted into EXPECTED_* in the test.
"""

import json