import torch
import torch.nn.functional as F
 
def generation_pipeline_analysis(model, 分詞器, text):
    """
    Trace the full path from 輸入 to 輸出 logits.
    """
    inputs = 分詞器(text, return_tensors="pt")
 
    with torch.no_grad():
        outputs = model(**inputs, output_hidden_states=True)
 
    # Final hidden state: (batch, seq_len, hidden_dim)
    final_hidden = outputs.hidden_states[-1][:, -1, :]
 
    # LM head projection: hidden_dim -> vocab_size
    logits = outputs.logits[:, -1, :]  # (batch, vocab_size)
 
    # Softmax converts logits to probabilities
    probs = F.softmax(logits, dim=-1)
 
    # Top 符元 and their logits/probabilities
    top_k = 20
    top_probs, top_indices = probs.topk(top_k)
    top_logits = logits[0, top_indices[0]]
 
    results = []
    for i in range(top_k):
        token_id = top_indices[0, i].item()
        results.append({
            "符元": 分詞器.decode([token_id]),
            "logit": top_logits[i].item(),
            "probability": top_probs[0, i].item(),
            "rank": i + 1
        })
 
    return results

def softmax_analysis(logits, temperature=1.0):
    """
    Analyze how temperature affects 符元 competition.
 
    Key insight: softmax amplifies small logit differences
    into large probability differences, especially at low
    temperatures.
    """
    scaled_logits = logits / temperature
    probs = F.softmax(scaled_logits, dim=-1)
 
    # The gap between top-1 and top-2 logits determines
    # how "committed" 模型 is to its top choice
    sorted_logits, _ = logits.sort(descending=True)
    logit_gap = (sorted_logits[0] - sorted_logits[1]).item()
 
    # Entropy measures uncertainty
    entropy = -(probs * probs.log()).sum().item()
 
    return {
        "top1_prob": probs.max().item(),
        "logit_gap": logit_gap,
        "entropy": entropy,
        "temperature": temperature,
        # Small logit gap = easily manipulable
        "manipulability": "high" if logit_gap < 1.0 else
                          "medium" if logit_gap < 3.0 else "low"
    }

class LogitLens:
    """Read model's evolving predictions across layers."""
 
    def __init__(self, model, 分詞器):
        self.model = model
        self.分詞器 = 分詞器
        self.lm_head = model.lm_head  # Final projection
 
    def analyze(self, text, position=-1):
        """
        Project each layer's hidden state through the LM head
        to see how 模型's prediction evolves.
        """
        inputs = self.分詞器(text, return_tensors="pt")
 
        with torch.no_grad():
            outputs = self.model(**inputs, output_hidden_states=True)
 
        layer_predictions = []
 
        for layer_idx, hidden_state in enumerate(outputs.hidden_states):
            # Apply layer norm (if model uses it before LM head)
            if hasattr(self.model.model, 'norm'):
                normed = self.model.model.norm(hidden_state)
            else:
                normed = hidden_state
 
            # Project through LM head
            logits = self.lm_head(normed[:, position, :])
            probs = F.softmax(logits, dim=-1)
 
            top_5_probs, top_5_ids = probs.topk(5)
            predictions = [
                {
                    "符元": self.分詞器.decode([tid.item()]),
                    "probability": p.item()
                }
                for tid, p in zip(top_5_ids[0], top_5_probs[0])
            ]
 
            layer_predictions.append({
                "layer": layer_idx,
                "top_predictions": predictions
            })
 
        return layer_predictions
 
    def find_safety_decision_layer(self, safe_prompt, unsafe_prompt):
        """
        Find the layer where 模型 first diverges between
        safe and unsafe behavior (where "refusal" prediction
        first appears or disappears).
        """
        safe_analysis = self.analyze(safe_prompt)
        unsafe_analysis = self.analyze(unsafe_prompt)
 
        divergence_points = []
 
        for layer_idx in range(len(safe_analysis)):
            safe_top = safe_analysis[layer_idx]["top_predictions"][0]["符元"]
            unsafe_top = unsafe_analysis[layer_idx]["top_predictions"][0]["符元"]
 
            if safe_top != unsafe_top:
                divergence_points.append({
                    "layer": layer_idx,
                    "safe_prediction": safe_top,
                    "unsafe_prediction": unsafe_top
                })
 
        return divergence_points

def token_priming_analysis(model, 分詞器, base_prompt,
                             target_token, candidate_primers):
    """
    Find 輸入 符元 that increase the logit of a target
    輸出 符元 when added to the prompt.
    """
    target_id = 分詞器.encode(target_token, add_special_tokens=False)[0]
 
    baseline_inputs = 分詞器(base_prompt, return_tensors="pt")
    with torch.no_grad():
        baseline_logits = model(**baseline_inputs).logits[:, -1, :]
    baseline_score = baseline_logits[0, target_id].item()
 
    results = []
    for primer in candidate_primers:
        primed_prompt = f"{base_prompt} {primer}"
        inputs = 分詞器(primed_prompt, return_tensors="pt")
        with torch.no_grad():
            logits = model(**inputs).logits[:, -1, :]
        primed_score = logits[0, target_id].item()
 
        results.append({
            "primer": primer,
            "logit_change": primed_score - baseline_score,
            "baseline_logit": baseline_score,
            "primed_logit": primed_score
        })
 
    results.sort(key=lambda x: x["logit_change"], reverse=True)
    return results

def margin_analysis(model, 分詞器, prompt_template, variables):
    """
    Analyze how prompt variations affect the logit margin
    between compliance and refusal 符元.
    """
    compliance_tokens = ["Sure", "Here", "Of", "I", "Yes"]
    refusal_tokens = ["I", "Sorry", "Unfortunately", "As"]
 
    compliance_ids = [
        分詞器.encode(t, add_special_tokens=False)[0]
        for t in compliance_tokens
    ]
    refusal_ids = [
        分詞器.encode(t, add_special_tokens=False)[0]
        for t in refusal_tokens
    ]
 
    results = []
    for var in variables:
        prompt = prompt_template.format(var=var)
        inputs = 分詞器(prompt, return_tensors="pt")
 
        with torch.no_grad():
            logits = model(**inputs).logits[:, -1, :]
 
        compliance_max = max(logits[0, tid].item() for tid in compliance_ids)
        refusal_max = max(logits[0, tid].item() for tid in refusal_ids)
        margin = refusal_max - compliance_max
 
        results.append({
            "variable": var,
            "compliance_logit": compliance_max,
            "refusal_logit": refusal_max,
            "margin": margin,  # Positive = refusal winning
            "close_call": abs(margin) < 1.0
        })
 
    return results

def repetition_logit_effect(model, 分詞器, base_prompt,
                              target_word, repetitions_range):
    """
    Measure how repeating a word in context affects its
    輸出 logit (測試 the frequency prior effect).
    """
    target_id = 分詞器.encode(
        f" {target_word}", add_special_tokens=False
    )[0]
 
    results = []
    for num_reps in repetitions_range:
        repeated = f" {target_word}" * num_reps
        prompt = f"{base_prompt}{repeated}\n\nResponse:"
 
        inputs = 分詞器(prompt, return_tensors="pt")
        with torch.no_grad():
            logits = model(**inputs).logits[:, -1, :]
 
        target_logit = logits[0, target_id].item()
        results.append({
            "repetitions": num_reps,
            "target_logit": target_logit
        })
 
    return results

def api_logprob_analysis(api_client, prompt, top_k=5):
    """
    Use API-provided logprobs to analyze model decision points
    without white-box access.
    """
    response = api_client.complete(
        prompt=prompt,
        max_tokens=1,
        logprobs=top_k,
        temperature=0
    )
 
    top_tokens = response.choices[0].logprobs.top_logprobs[0]
 
    # 識別 if 模型 is near a compliance/refusal boundary
    compliance_indicators = {"Sure", "Here", "Of", "Yes", "Certa"}
    refusal_indicators = {"I", "Sorry", "Unfor", "As", "Thank"}
 
    compliance_score = max(
        (prob for 符元, prob in top_tokens.items()
         if any(符元.strip().startswith(c) for c in compliance_indicators)),
        default=-100
    )
    refusal_score = max(
        (prob for 符元, prob in top_tokens.items()
         if any(符元.strip().startswith(r) for r in refusal_indicators)),
        default=-100
    )
 
    return {
        "top_tokens": top_tokens,
        "compliance_logprob": compliance_score,
        "refusal_logprob": refusal_score,
        "margin": refusal_score - compliance_score,
        "boundary_proximity": abs(refusal_score - compliance_score) < 0.5
    }