Claude commited on
Commit
ea6878f
·
1 Parent(s): 37d0b8b

Fix AC3: Switch from Llama 3.1 70B to Llama 3.2 3B Instruct

Browse files

Llama 3.1 70B (70 billion parameters) is too large for HuggingFace's
free Serverless Inference API (limit ~10GB). This was causing 404 errors.

Changes:
- Updated model from "meta-llama/Llama-3.1-70B-Instruct" to "meta-llama/Llama-3.2-3B-Instruct"
- Llama 3.2 3B Instruct is available for FREE on HF Inference API
- Updated documentation comments to reflect the new model
- Model is still capable of generating JSON for UI variants

This should fix the 404 errors and enable AI-powered variant generation.

Files changed (1) hide show
  1. agents/flutter_screen_generator.py +4 -4
agents/flutter_screen_generator.py CHANGED
@@ -319,7 +319,7 @@ class FlutterScreenGenerator:
319
  self.hf_token = os.getenv("HF_TOKEN", "")
320
  # Explicitly use HF provider (not third-party like Hyperbolic)
321
  self.client = InferenceClient(token=self.hf_token, provider="hf-inference")
322
- self.model = "meta-llama/Llama-3.1-70B-Instruct"
323
 
324
  # Jinja2 setup with custom delimiters (to avoid conflict with Dart ${})
325
  template_dir = os.path.join(os.path.dirname(__file__), '..', 'templates', 'dart')
@@ -463,7 +463,7 @@ class FlutterScreenGenerator:
463
 
464
  def _generate_plan(self, theme: Dict[str, Any], app_type: str, screen_type: str) -> Dict[str, Any]:
465
  """
466
- Use Llama 3.1 70B to generate architectural plan for the screen
467
 
468
  Returns JSON plan with sections, components, and layout decisions
469
  """
@@ -471,7 +471,7 @@ class FlutterScreenGenerator:
471
  prompt = self._build_planning_prompt(theme, app_type, screen_type)
472
 
473
  try:
474
- # Call Llama 3.1 70B via HuggingFace Inference API
475
  generated_text = self.client.text_generation(
476
  prompt=prompt,
477
  model=self.model,
@@ -542,7 +542,7 @@ JSON Plan:"""
542
 
543
  def _generate_variants_with_ai(self, theme: Dict[str, Any], app_type: str) -> list:
544
  """
545
- Use Llama 3.1 70B to generate 4 intelligent UI variants
546
  Each variant combines background + header + footer appropriately
547
  """
548
 
 
319
  self.hf_token = os.getenv("HF_TOKEN", "")
320
  # Explicitly use HF provider (not third-party like Hyperbolic)
321
  self.client = InferenceClient(token=self.hf_token, provider="hf-inference")
322
+ self.model = "meta-llama/Llama-3.2-3B-Instruct" # Free tier model (70B too large for serverless API)
323
 
324
  # Jinja2 setup with custom delimiters (to avoid conflict with Dart ${})
325
  template_dir = os.path.join(os.path.dirname(__file__), '..', 'templates', 'dart')
 
463
 
464
  def _generate_plan(self, theme: Dict[str, Any], app_type: str, screen_type: str) -> Dict[str, Any]:
465
  """
466
+ Use Llama 3.2 3B Instruct to generate architectural plan for the screen
467
 
468
  Returns JSON plan with sections, components, and layout decisions
469
  """
 
471
  prompt = self._build_planning_prompt(theme, app_type, screen_type)
472
 
473
  try:
474
+ # Call Llama 3.2 3B Instruct via HuggingFace Inference API
475
  generated_text = self.client.text_generation(
476
  prompt=prompt,
477
  model=self.model,
 
542
 
543
  def _generate_variants_with_ai(self, theme: Dict[str, Any], app_type: str) -> list:
544
  """
545
+ Use Llama 3.2 3B Instruct to generate 4 intelligent UI variants
546
  Each variant combines background + header + footer appropriately
547
  """
548