josephrw commited on
Commit
d39cb8d
·
verified ·
1 Parent(s): 3c5ff56

Upload folder using huggingface_hub

Browse files
Files changed (2) hide show
  1. src/builder_llm.py +1 -1
  2. src/observer_llm.py +9 -26
src/builder_llm.py CHANGED
@@ -57,7 +57,7 @@ def call_hf_inference(prompt: str) -> str:
57
 
58
  token = os.getenv("HF_TOKEN") or os.getenv("HUGGING_FACE_HUB_TOKEN")
59
  client = InferenceClient(token=token)
60
- model = os.getenv("HF_TEXT_MODEL", "meta-llama/Llama-3.1-8B-Instruct")
61
  response = client.chat_completion(
62
  model=model,
63
  messages=[{"role": "user", "content": prompt}],
 
57
 
58
  token = os.getenv("HF_TOKEN") or os.getenv("HUGGING_FACE_HUB_TOKEN")
59
  client = InferenceClient(token=token)
60
+ model = os.getenv("HF_TEXT_MODEL", "Qwen/Qwen2.5-7B-Instruct")
61
  response = client.chat_completion(
62
  model=model,
63
  messages=[{"role": "user", "content": prompt}],
src/observer_llm.py CHANGED
@@ -63,37 +63,20 @@ def call_openai(prompt: str, frame_b64: str | None = None) -> str:
63
  def call_hf_inference(prompt: str, frame_b64: str | None = None) -> str:
64
  """Call HF Inference API using huggingface_hub InferenceClient.
65
 
66
- Uses vision-language model when frame is available.
67
- Falls back to text-only model for state-only observations.
68
  """
69
  from huggingface_hub import InferenceClient
70
 
71
  token = os.getenv("HF_TOKEN") or os.getenv("HUGGING_FACE_HUB_TOKEN")
72
  client = InferenceClient(token=token)
73
-
74
- if frame_b64:
75
- model = os.getenv("HF_VISION_MODEL", "meta-llama/Llama-3.2-11B-Vision-Instruct")
76
- image_url = f"data:image/jpeg;base64,{frame_b64}"
77
- response = client.chat_completion(
78
- model=model,
79
- messages=[{
80
- "role": "user",
81
- "content": [
82
- {"type": "text", "text": prompt},
83
- {"type": "image_url", "image_url": {"url": image_url}},
84
- ],
85
- }],
86
- max_tokens=800,
87
- )
88
- return response.choices[0].message.content
89
- else:
90
- model = os.getenv("HF_TEXT_MODEL", "meta-llama/Llama-3.1-8B-Instruct")
91
- response = client.chat_completion(
92
- model=model,
93
- messages=[{"role": "user", "content": prompt}],
94
- max_tokens=800,
95
- )
96
- return response.choices[0].message.content
97
 
98
 
99
  def observe(state: SessionState, compact_state: dict) -> dict:
 
63
  def call_hf_inference(prompt: str, frame_b64: str | None = None) -> str:
64
  """Call HF Inference API using huggingface_hub InferenceClient.
65
 
66
+ Uses text model with compressed sensory state as text input.
67
+ Vision encoding (entropy, motion, phash) is already done by vision_sampler.
68
  """
69
  from huggingface_hub import InferenceClient
70
 
71
  token = os.getenv("HF_TOKEN") or os.getenv("HUGGING_FACE_HUB_TOKEN")
72
  client = InferenceClient(token=token)
73
+ model = os.getenv("HF_TEXT_MODEL", "Qwen/Qwen2.5-7B-Instruct")
74
+ response = client.chat_completion(
75
+ model=model,
76
+ messages=[{"role": "user", "content": prompt}],
77
+ max_tokens=800,
78
+ )
79
+ return response.choices[0].message.content
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
80
 
81
 
82
  def observe(state: SessionState, compact_state: dict) -> dict: