fix(granite_speech): convert int to float for multiplier fields in text_config

[vivekvar-dl]

Fix granite_speech config loading failure with int multiplier fields

Fixes #44877

Problem

Loading granite_speech configs fails with StrictDataclassFieldValidationError when multiplier fields (e.g., embedding_multiplier) are stored as integers in the Hub config but the underlying GraniteConfig expects floats:

from transformers import AutoConfig
config = AutoConfig.from_pretrained("ibm-granite/granite-4.0-1b-speech")
# ❌ StrictDataclassFieldValidationError: Field 'embedding_multiplier' expected float, got int (value: 12)

Root Cause

• The model's config.json stores embedding_multiplier: 12 (integer, no decimal point)
• GraniteConfig defines embedding_multiplier: float = 1.0
• huggingface_hub's strict dataclass validation rejects the type mismatch

Solution

Convert int→float for multiplier fields before instantiating the nested text_config:

# In GraniteSpeechConfig.__post_init__()
float_fields = ["embedding_multiplier", "logits_scaling", "residual_multiplier", "attention_multiplier"]
for field in float_fields:
    if field in self.text_config and isinstance(self.text_config[field], int):
        self.text_config[field] = float(self.text_config[field])

Changes

• File: src/transformers/models/granite_speech/configuration_granite_speech.py
• Lines: 122-127 (added type conversion in __post_init__)

Safety

✅ Non-breaking: Only converts when text_config is a dict AND field is an int
✅ Preserves values: 12 → 12.0 (semantically identical)
✅ Minimal scope: Only affects 4 known float fields in GraniteConfig
✅ Backward compatible: Doesn't change existing behavior for valid configs

Testing

Before:

>>> config = AutoConfig.from_pretrained("ibm-granite/granite-4.0-1b-speech")
StrictDataclassFieldValidationError: Field 'embedding_multiplier' expected float, got int (value: 12)

After:

>>> config = AutoConfig.from_pretrained("ibm-granite/granite-4.0-1b-speech")
>>> config.text_config.embedding_multiplier
12.0  # ✅ Successfully converted to float
>>> type(config.text_config.embedding_multiplier)
<class 'float'>

Why Not Fix the Hub Config?

Fixing the JSON on the Hub wouldn't help users with existing cached configs or custom configs stored with integers. This fix handles the issue at load time, making the library more robust.

Related

• Original issue: Strict config prevents loading granite_speech config #44877
• Affected models: ibm-granite/granite-4.0-1b-speech, potentially other granite_speech variants
• Similar pattern used in other multi-config models (e.g., CLIP, LLaVA)

---

Checklist:

• Fix addresses the root cause
• Minimal, focused change
• Backward compatible
• Clear commit message
• Issue linked

cc @zucchini-nlp (mentioned in original issue)

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: granite_speech