Revert "Add Qwen 0.8 3.5"

This reverts commit 762b54a0a7c471930f8a8d966027031e1d7097f0.

README.md

@@ -11,7 +11,4 @@ license: mit
 short_description: Steer LLMs with activation vectors
 ---
 
-An interactive Gradio demo that lets you steer local models using activation vectors.
-Included model options: `LiquidAI/LFM2-350M`, `LiquidAI/LFM2-700M`, `LiquidAI/LFM2-1.2B`, and `Qwen/Qwen3.5-0.8B`.
-
-For `Qwen/Qwen3.5-0.8B`, use a recent `transformers` version that recognizes the `qwen3_5` architecture.
+An interactive Gradio demo that lets you steer two local models using activation vectors on layer 8. Models: LiquidAI/LFM2-350M and unsloth/Llama-3.2-1B-Instruct. Add one or more vectors with coefficients, pick the model, and chat with the steered output.

app.py

@@ -8,12 +8,7 @@ import torch
 from llm_steer import Steer, DecaySchedule
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
-MODEL_IDS = [
-    "LiquidAI/LFM2-350M",
-    "LiquidAI/LFM2-700M",
-    "LiquidAI/LFM2-1.2B",
-    "Qwen/Qwen3.5-0.8B",
-]
+MODEL_IDS = ["LiquidAI/LFM2-350M", "LiquidAI/LFM2-700M", "LiquidAI/LFM2-1.2B"]
 DEFAULT_SCHEDULE_ENABLED = True
 DEFAULT_SCHEDULE_RATE = 0.85
 DEFAULT_SCHEDULE_MIN_MULTIPLIER = 0.35
@@ -96,34 +91,7 @@ def _load_model(model_id: str) -> ModelBundle:
     )
 
 
-MODELS: dict[str, ModelBundle] = {}
-MODEL_LOAD_LOCK = threading.Lock()
-
-
-def _get_model_bundle(model_id: str) -> ModelBundle:
-    bundle = MODELS.get(model_id)
-    if bundle is not None:
-        return bundle
-    with MODEL_LOAD_LOCK:
-        bundle = MODELS.get(model_id)
-        if bundle is None:
-            bundle = _load_model(model_id)
-            MODELS[model_id] = bundle
-    return bundle
-
-
-def _format_model_load_error(model_id: str, exc: Exception) -> str:
-    reason = str(exc).strip()
-    if "does not recognize this architecture" in reason:
-        return (
-            f"Could not load {model_id}. "
-            "Your installed `transformers` version is too old for this model. "
-            "Upgrade `transformers` and retry."
-        )
-    if reason:
-        first_line = reason.splitlines()[0].strip()
-        return f"Could not load {model_id}: {first_line}"
-    return f"Could not load {model_id}."
+MODELS = {model_id: _load_model(model_id) for model_id in MODEL_IDS}
 
 
 def _parse_vectors(
@@ -255,15 +223,7 @@ def respond(
     if not message:
         return history, ""
 
-    try:
-        bundle = _get_model_bundle(model_id)
-    except Exception as exc:
-        updated_history = history + [
-            {"role": "user", "content": message},
-            {"role": "assistant", "content": _format_model_load_error(model_id, exc)},
-        ]
-        return updated_history, ""
-
+    bundle = MODELS[model_id]
     steer_vectors = _parse_vectors(vectors)
     schedule = None
     if vectors_enabled and schedule_enabled:
@@ -417,7 +377,7 @@ with gr.Blocks(theme=THEME, css=CSS) as demo:
                 )
                 add_vector = gr.Button("Add new vector row")
                 gr.Markdown(
-                    "Coeff can be negative. Layer index should match model depth (LFM2: 0-15, Qwen3.5-0.8B: 0-23)."
+                    "Coeff can be negative. Layer should have value between 0 and 15."
                 )
             with gr.Accordion(
                 "Steering schedule",

requirements.txt

@@ -1,5 +1,5 @@
 llm_steer >= 2.0.1
-transformers>=5.2.0
+transformers
 torch
 accelerate
 sentencepiece