Auto upload zain 2026-08-12T20:48:44.955825

zain/Activation/README.md

@@ -0,0 +1 @@
+# Activation

zain/Activation/exp.py

@@ -0,0 +1,1102 @@
+"""
+Tiny Llama GLU Lab — Consolidated training library.
+One file: model definition, activation registry, stability monitoring,
+time tracking, dataset builder, and Trainer factory.
+"""
+
+import math
+import os
+import time
+import json
+import re
+from pathlib import Path
+from itertools import chain
+from typing import Dict, Callable, Optional, List, Any, Tuple
+
+import torch
+import torch.nn as nn
+from transformers import (
+    LlamaConfig,
+    LlamaPreTrainedModel,
+    Trainer,
+    TrainerCallback,
+    TrainingArguments,
+    DataCollatorForLanguageModeling,
+    AutoTokenizer,
+    set_seed,
+)
+from transformers.models.llama.modeling_llama import (
+    LlamaAttention,
+    LlamaRMSNorm,
+    LlamaRotaryEmbedding,
+)
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from datasets import load_dataset
+
+
+# =============================================================================
+# 1. ACTIVATION REGISTRY
+# =============================================================================
+
+class GLUActivationRegistry:
+    """Own every gating activation you test. Add new variants in one line."""
+    _registry: Dict[str, Callable[[torch.Tensor], torch.Tensor]] = {}
+
+    @classmethod
+    def register(cls, name: str, fn: Callable[[torch.Tensor], torch.Tensor]) -> None:
+        cls._registry[name] = fn
+
+    @classmethod
+    def get(cls, name: str) -> Callable[[torch.Tensor], torch.Tensor]:
+        if name not in cls._registry:
+            raise KeyError(
+                f"Activation '{name}' not found. Available: {list(cls._registry.keys())}"
+            )
+        return cls._registry[name]
+
+
+# Built-ins
+GLUActivationRegistry.register("silu", nn.functional.silu)
+GLUActivationRegistry.register("swish", nn.functional.silu)
+GLUActivationRegistry.register("relu", nn.functional.relu)
+GLUActivationRegistry.register("gelu", nn.functional.gelu)
+GLUActivationRegistry.register("sigmoid", torch.sigmoid)
+GLUActivationRegistry.register("tanh", torch.tanh)
+GLUActivationRegistry.register("softplus", nn.functional.softplus)
+GLUActivationRegistry.register("linear", lambda x: x)
+
+# Custom
+GLUActivationRegistry.register("s10", lambda x: x * x * torch.sigmoid(x))
+GLUActivationRegistry.register("w1a", lambda x: x * torch.tanh(x))
+
+
+# =============================================================================
+# 2. CONFIG
+# =============================================================================
+
+class TinyLlamaConfig(LlamaConfig):
+    """
+    Exact Llama config plus two fields:
+      - mlp_type: "glu" or "mlp" (standard Transformer MLP)
+      - activation: name of the activation function to use inside the MLP block.
+      - waleed_beta: β_cap for the waleed10 post‑clip (default 10.0)
+    Enforces pure MHA by requiring num_key_value_heads == num_attention_heads.
+    """
+    model_type = "tiny_llama"
+
+    def __init__(
+        self,
+        mlp_type: str = "glu",
+        activation: str = "silu",
+        waleed_beta: float = 10.0,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+        self.mlp_type = mlp_type
+        self.activation = activation
+        self.waleed_beta = waleed_beta
+        if self.num_key_value_heads != self.num_attention_heads:
+            raise ValueError(
+                f"Pure MHA required: num_key_value_heads ({self.num_key_value_heads}) "
+                f"must equal num_attention_heads ({self.num_attention_heads})."
+            )
+
+
+# =============================================================================
+# 3. MODEL
+# =============================================================================
+
+class TinyLlamaMLP(nn.Module):
+    """
+    Unified MLP block supporting:
+      - Standard GLU:   down_proj( act(gate_proj(x)) * up_proj(x) )
+      - Standard MLP:   down_proj( act(up_proj(x)) )
+      - Gated variants (situglu, waleed) with internal tanh scaling.
+      - waleed10 / silu-waleed10 with post‑down‑proj clipping.
+
+    All projections are always named the same way, ensuring consistent
+    tensor logging regardless of activation.
+    """
+    def __init__(self, config: TinyLlamaConfig):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.mlp_type = config.mlp_type
+        self.activation_name = config.activation
+        self.waleed_beta = getattr(config, "waleed_beta", 10.0)
+
+        # Effective intermediate size (scaled for MLP)
+        if self.mlp_type == "glu":
+            effective_intermediate = self.intermediate_size
+        elif self.mlp_type == "mlp":
+            effective_intermediate = int(self.intermediate_size * 1.5)
+            print(f"[MLP] Auto‑scaled intermediate_size from {self.intermediate_size} to {effective_intermediate} for parameter parity.")
+        else:
+            raise ValueError(f"Unknown mlp_type: {self.mlp_type}")
+
+        self.effective_intermediate = effective_intermediate
+
+        # Always define the projections (names consistent across variants)
+        if self.mlp_type == "glu":
+            self.gate_proj = nn.Linear(self.hidden_size, effective_intermediate, bias=False)
+            self.up_proj = nn.Linear(self.hidden_size, effective_intermediate, bias=False)
+        else:  # mlp
+            self.up_proj = nn.Linear(self.hidden_size, effective_intermediate, bias=False)
+
+        self.down_proj = nn.Linear(effective_intermediate, self.hidden_size, bias=False)
+
+        # Activation function (for standard variants)
+        if self.mlp_type == "glu":
+            if self.activation_name in ("situglu", "waleed", "situglu_low", "waleedglu_low"):
+                self.act_fn = None
+            elif self.activation_name == "waleed10":
+                self.act_fn = GLUActivationRegistry.get("linear")
+            elif self.activation_name == "silu-waleed10":
+                self.act_fn = GLUActivationRegistry.get("silu")
+            else:
+                self.act_fn = GLUActivationRegistry.get(self.activation_name)
+        else:  # mlp
+            if self.activation_name in ("situglu", "waleed", "situglu_low", "waleedglu_low"):
+                raise ValueError(
+                    f"Activation '{self.activation_name}' requires a gated architecture (GLU). "
+                    f"Please use mlp_type='glu'."
+                )
+            elif self.activation_name == "waleed10":
+                self.act_fn = GLUActivationRegistry.get("linear")
+            elif self.activation_name == "silu-waleed10":
+                self.act_fn = GLUActivationRegistry.get("silu")
+            else:
+                self.act_fn = GLUActivationRegistry.get(self.activation_name)
+
+        # Determine beta values for situglu/waleed variants
+        if self.activation_name in ("situglu_low", "waleedglu_low"):
+            self.beta1 = 2.5
+            self.beta2 = 4.0
+        else:
+            self.beta1 = 4.0
+            self.beta2 = 25.0
+
+        # Flags for special handling
+        self.is_situglu = self.activation_name in ("situglu", "situglu_low")
+        self.is_waleed = self.activation_name in ("waleed", "waleedglu_low")
+        self.is_waleed10 = self.activation_name in ("waleed10", "silu-waleed10")
+        self.has_sigmoid_gate = self.activation_name.startswith("situglu")
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if self.mlp_type == "glu":
+            gate = self.gate_proj(x)
+            up = self.up_proj(x)
+
+            if self.is_situglu or self.is_waleed:
+                if self.has_sigmoid_gate:
+                    gate = self.beta1 * torch.tanh(gate / self.beta1) * torch.sigmoid(gate)
+                else:
+                    gate = self.beta1 * torch.tanh(gate / self.beta1)
+                up = self.beta2 * torch.tanh(up / self.beta2)
+                hidden = gate * up
+            else:
+                hidden = self.act_fn(gate) * up
+
+            out = self.down_proj(hidden)
+
+        else:  # mlp
+            hidden = self.act_fn(self.up_proj(x))
+            out = self.down_proj(hidden)
+
+        if self.is_waleed10:
+            out = self.waleed_beta * torch.tanh(out / self.waleed_beta)
+
+        return out
+
+
+class TinyLlamaDecoderLayer(nn.Module):
+    def __init__(self, config: TinyLlamaConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = LlamaAttention(config=config, layer_idx=layer_idx)
+        self.mlp = TinyLlamaMLP(config)
+        self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = LlamaRMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+        # Residual stream gateways (zero params, hookable)
+        self.residual_pre_attn = nn.Identity()
+        self.residual_post_attn = nn.Identity()
+        self.residual_post_mlp = nn.Identity()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        **kwargs,
+    ):
+        # Attention sub-layer
+        residual = hidden_states
+        hidden_states = self.residual_pre_attn(hidden_states)
+        hidden_states = self.input_layernorm(hidden_states)
+        attn_out = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            position_embeddings=position_embeddings,
+        )[0]
+        hidden_states = residual + attn_out
+        hidden_states = self.residual_post_attn(hidden_states)
+
+        # MLP sub-layer
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+        hidden_states = self.residual_post_mlp(hidden_states)
+        return (hidden_states,)
+
+
+# ----------------------------------------------------------------------------
+# ATTENTION MASK
+# ----------------------------------------------------------------------------
+def _build_causal_mask(
+    attention_mask: Optional[torch.Tensor],
+    seq_len: int,
+    dtype: torch.dtype,
+    device: torch.device,
+) -> torch.Tensor:
+    min_value = torch.finfo(dtype).min
+    causal = torch.full((seq_len, seq_len), fill_value=min_value, dtype=dtype, device=device)
+    causal = torch.triu(causal, diagonal=1)
+    causal = causal[None, None, :, :]
+
+    if attention_mask is None:
+        batch_size = 1
+        return causal.expand(batch_size, 1, seq_len, seq_len)
+
+    batch_size = attention_mask.shape[0]
+    causal = causal.expand(batch_size, 1, seq_len, seq_len).clone()
+    padding = attention_mask[:, None, None, :].to(device) == 0
+    causal = causal.masked_fill(padding, min_value)
+    return causal
+
+
+_MASK_PRINTED = False
+
+
+class TinyLlamaModel(LlamaPreTrainedModel):
+    config_class = TinyLlamaConfig
+
+    def __init__(self, config: TinyLlamaConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+        self.embed_tokens = nn.Embedding(
+            config.vocab_size, config.hidden_size, self.padding_idx
+        )
+        self.layers = nn.ModuleList(
+            [TinyLlamaDecoderLayer(config, i) for i in range(config.num_hidden_layers)]
+        )
+        self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = LlamaRotaryEmbedding(config=config)
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ):
+        global _MASK_PRINTED
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if position_ids is None:
+            seq_len = inputs_embeds.shape[1]
+            position_ids = torch.arange(
+                seq_len, device=inputs_embeds.device
+            ).unsqueeze(0).expand(inputs_embeds.shape[0], -1)
+
+        hidden_states = inputs_embeds
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        seq_len = hidden_states.shape[1]
+        causal_mask = _build_causal_mask(
+            attention_mask, seq_len, hidden_states.dtype, hidden_states.device
+        )
+
+        if not _MASK_PRINTED:
+            print("[INFO] Causal mask (float with -inf) applied to all attention layers.")
+            _MASK_PRINTED = True
+
+        for decoder_layer in self.layers:
+            layer_outputs = decoder_layer(
+                hidden_states,
+                attention_mask=causal_mask,
+                position_ids=position_ids,
+                position_embeddings=position_embeddings,
+            )
+            hidden_states = layer_outputs[0]
+
+        hidden_states = self.norm(hidden_states)
+        if not return_dict:
+            return (hidden_states,)
+        return {"last_hidden_state": hidden_states, "hidden_states": None, "attentions": None}
+
+
+class TinyLlamaForCausalLM(LlamaPreTrainedModel):
+    config_class = TinyLlamaConfig
+    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
+
+    def __init__(self, config: TinyLlamaConfig):
+        super().__init__(config)
+        self.model = TinyLlamaModel(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        if config.tie_word_embeddings:
+            self.lm_head.weight = self.model.embed_tokens.weight
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ):
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            return_dict=return_dict,
+        )
+        hidden_states = outputs["last_hidden_state"] if return_dict else outputs[0]
+        logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            loss_fct = nn.CrossEntropyLoss()
+            loss = loss_fct(
+                shift_logits.view(-1, self.config.vocab_size), shift_labels.view(-1)
+            )
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=None,
+            hidden_states=None,
+            attentions=None,
+        )
+
+    def prepare_inputs_for_generation(
+        self, input_ids, past_key_values=None, attention_mask=None, **kwargs
+    ):
+        if past_key_values:
+            input_ids = input_ids[:, -1:]
+        position_ids = kwargs.get("position_ids")
+        if attention_mask is not None and position_ids is None:
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -1].unsqueeze(-1)
+        return {
+            "input_ids": input_ids,
+            "position_ids": position_ids,
+            "past_key_values": past_key_values,
+            "attention_mask": attention_mask,
+        }
+
+
+# =============================================================================
+# 4. MONITORING ENGINE
+# =============================================================================
+
+class StatsEngine:
+    """Compute unified signature for any tensor."""
+
+    @staticmethod
+    def compute(
+        tensor: torch.Tensor, user_limit: float, dtype_ratio: float
+    ) -> Dict[str, float]:
+        with torch.no_grad():
+            abs_t = tensor.abs()
+            dtype_info = torch.finfo(tensor.dtype)
+            dtype_limit = (
+                dtype_ratio * dtype_info.max
+                if not torch.isinf(torch.tensor(dtype_info.max))
+                else float("inf")
+            )
+
+            t_min = tensor.min().item()
+            t_max = tensor.max().item()
+
+            return {
+                "norm": tensor.norm(2).item(),
+                "mean": tensor.mean().item(),
+                "std": tensor.std().item(),
+                "max_abs": abs_t.max().item(),
+                "frac_near_dtype_limit": (
+                    (abs_t > dtype_limit).float().mean().item()
+                    if not math.isinf(dtype_limit)
+                    else 0.0
+                ),
+                "frac_near_user_limit": (abs_t > user_limit).float().mean().item(),
+                "min": t_min,
+                "max": t_max,
+                "range": t_max - t_min,
+            }
+
+
+class StepAccumulator:
+    """Stores per-tensor entries, aggregates to layer/global scope."""
+
+    def __init__(self):
+        self.tensors: Dict[str, Dict[str, float]] = {}
+
+    def add(self, name: str, numel: int, stats: Dict[str, float]):
+        new_entry = {"numel": numel, **stats}
+        existing = self.tensors.get(name)
+        self.tensors[name] = (
+            new_entry if existing is None else self._merge_entry(existing, new_entry)
+        )
+
+    @staticmethod
+    def _merge_entry(a: Dict[str, float], b: Dict[str, float]) -> Dict[str, float]:
+        total_n = a["numel"] + b["numel"]
+        if total_n == 0:
+            return a
+        norm = math.sqrt(a["norm"] ** 2 + b["norm"] ** 2)
+        max_abs = max(a["max_abs"], b["max_abs"])
+        mean = (a["mean"] * a["numel"] + b["mean"] * b["numel"]) / total_n
+        ex2 = (
+            a["numel"] * (a["std"] ** 2 + a["mean"] ** 2)
+            + b["numel"] * (b["std"] ** 2 + b["mean"] ** 2)
+        ) / total_n
+        std = math.sqrt(max(0.0, ex2 - mean ** 2))
+        frac_dtype = (
+            a["frac_near_dtype_limit"] * a["numel"] + b["frac_near_dtype_limit"] * b["numel"]
+        ) / total_n
+        frac_user = (
+            a["frac_near_user_limit"] * a["numel"] + b["frac_near_user_limit"] * b["numel"]
+        ) / total_n
+
+        t_min = min(a.get("min", float("inf")), b.get("min", float("inf")))
+        t_max = max(a.get("max", float("-inf")), b.get("max", float("-inf")))
+
+        return {
+            "numel": total_n,
+            "norm": norm,
+            "mean": mean,
+            "std": std,
+            "max_abs": max_abs,
+            "frac_near_dtype_limit": frac_dtype,
+            "frac_near_user_limit": frac_user,
+            "min": t_min,
+            "max": t_max,
+            "range": t_max - t_min,
+        }
+
+    def clear(self):
+        self.tensors.clear()
+
+    def _aggregate(self, entries: Dict[str, Dict[str, float]]) -> Dict[str, float]:
+        if not entries:
+            return {}
+        numels = [e["numel"] for e in entries.values()]
+        total_n = sum(numels)
+
+        norm = math.sqrt(sum(e["norm"] ** 2 for e in entries.values()))
+        max_abs = max(e["max_abs"] for e in entries.values())
+        mean = sum(e["mean"] * e["numel"] for e in entries.values()) / total_n
+        ex2 = (
+            sum(e["numel"] * (e["std"] ** 2 + e["mean"] ** 2) for e in entries.values())
+            / total_n
+        )
+        std = math.sqrt(max(0.0, ex2 - mean ** 2))
+        frac_dtype = (
+            sum(e["frac_near_dtype_limit"] * e["numel"] for e in entries.values())
+            / total_n
+        )
+        frac_user = (
+            sum(e["frac_near_user_limit"] * e["numel"] for e in entries.values())
+            / total_n
+        )
+
+        t_min = min(e.get("min", float("inf")) for e in entries.values())
+        t_max = max(e.get("max", float("-inf")) for e in entries.values())
+
+        return {
+            "norm": norm,
+            "mean": mean,
+            "std": std,
+            "max_abs": max_abs,
+            "frac_near_dtype_limit": frac_dtype,
+            "frac_near_user_limit": frac_user,
+            "min": t_min,
+            "max": t_max,
+            "range": t_max - t_min,
+        }
+
+    def get_global_stats(self) -> Dict[str, float]:
+        return self._aggregate(self.tensors)
+
+    def get_layer_stats(self, layer_prefix: str) -> Dict[str, float]:
+        entries = {
+            k: v for k, v in self.tensors.items() if k.startswith(layer_prefix + ".")
+        }
+        return self._aggregate(entries)
+
+
+class HookRegistry:
+    """Attach and throttle forward/backward hooks."""
+
+    def __init__(self, model: nn.Module):
+        self.model = model
+        self.handles: List[torch.utils.hooks.RemovableHandle] = []
+        self.active = False
+
+    def attach_forward(
+        self,
+        module_patterns: List[str],
+        accumulator: StepAccumulator,
+        user_limit: float,
+        dtype_ratio: float,
+    ):
+        for name, module in self.model.named_modules():
+            if any(re.search(p, name) for p in module_patterns):
+                h = module.register_forward_hook(
+                    self._make_forward_hook(name, accumulator, user_limit, dtype_ratio)
+                )
+                self.handles.append(h)
+
+    def attach_backward(
+        self, accumulator: StepAccumulator, user_limit: float, dtype_ratio: float
+    ):
+        for name, param in self.model.named_parameters():
+            if param.requires_grad:
+                h = param.register_hook(
+                    self._make_backward_hook(
+                        f"grad.{name}", accumulator, user_limit, dtype_ratio
+                    )
+                )
+                self.handles.append(h)
+
+    def _make_forward_hook(
+        self, module_name: str, accumulator: StepAccumulator, user_limit: float, dtype_ratio: float
+    ):
+        def hook(module, inp, out):
+            if not self.active:
+                return
+            if isinstance(out, dict):
+                out_dict = out
+                out = out_dict.get("last_hidden_state")
+                if out is None:
+                    out = next(
+                        (v for v in out_dict.values() if torch.is_tensor(v)), None
+                    )
+            elif isinstance(out, (tuple, list)):
+                out = out[0] if len(out) > 0 else None
+
+            if not torch.is_tensor(out):
+                return
+
+            stats = StatsEngine.compute(out.detach(), user_limit, dtype_ratio)
+            accumulator.add(f"act.{module_name}", out.numel(), stats)
+
+        return hook
+
+    def _make_backward_hook(
+        self, param_name: str, accumulator: StepAccumulator, user_limit: float, dtype_ratio: float
+    ):
+        def hook(grad):
+            if not self.active:
+                return
+            stats = StatsEngine.compute(grad.detach(), user_limit, dtype_ratio)
+            accumulator.add(param_name, grad.numel(), stats)
+
+        return hook
+
+    def set_active(self, active: bool):
+        self.active = active
+
+    def clear(self):
+        for h in self.handles:
+            h.remove()
+        self.handles.clear()
+
+
+class StabilityMonitorCallback(TrainerCallback):
+    """Full stability instrumentation: grad / param / act statistics."""
+
+    def __init__(
+        self,
+        model: nn.Module,
+        monitor_every_n_steps: int = 10,
+        module_patterns: Optional[List[str]] = None,
+        user_limits: Optional[Dict[str, float]] = None,
+        dtype_proximity_ratio: float = 0.9,
+        log_scope: Optional[Dict[str, bool]] = None,
+        monitor_during_eval: bool = False,
+    ):
+        self.model = model
+        self.monitor_every_n_steps = monitor_every_n_steps
+        self.module_patterns = module_patterns or [".*mlp.*", ".*self_attn.*", ".*residual.*"]
+        self.user_limits = user_limits or {"grad": 1.0, "param": 100.0, "act": 50.0}
+        self.dtype_ratio = dtype_proximity_ratio
+        self.log_scope = log_scope or {
+            "global": True,
+            "per_layer": True,
+            "per_tensor": False,
+        }
+        self.monitor_during_eval = monitor_during_eval
+
+        self.accumulator = StepAccumulator()
+        self.hooks = HookRegistry(model)
+        self.hooks.attach_forward(
+            self.module_patterns,
+            self.accumulator,
+            self.user_limits["act"],
+            self.dtype_ratio,
+        )
+        self.hooks.attach_backward(
+            self.accumulator, self.user_limits["grad"], self.dtype_ratio
+        )
+
+        self.pending_metrics: Optional[Dict[str, float]] = None
+
+    def _should_monitor(self, state) -> bool:
+        return state.global_step % self.monitor_every_n_steps == 0
+
+    def on_step_begin(self, args, state, control, **kwargs):
+        if self._should_monitor(state):
+            self.accumulator.clear()
+            self.hooks.set_active(True)
+
+    def on_step_end(self, args, state, control, **kwargs):
+        if not self.hooks.active:
+            return
+        for name, param in self.model.named_parameters():
+            stats = StatsEngine.compute(
+                param.data, self.user_limits["param"], self.dtype_ratio
+            )
+            self.accumulator.add(f"param.{name}", param.numel(), stats)
+
+        self.hooks.set_active(False)
+        self.pending_metrics = self._build_metrics()
+
+    @staticmethod
+    def _kind_of(name: str) -> str:
+        if name.startswith("act."):
+            return "act"
+        if name.startswith("grad."):
+            return "grad"
+        if name.startswith("param."):
+            return "param"
+        return "other"
+
+    @staticmethod
+    def _strip_kind(name: str) -> str:
+        if name.startswith("act."):
+            return name[4:]
+        if name.startswith(("grad.", "param.")):
+            return name[5:]
+        return name
+
+    def _build_metrics(self, scope: str = "train") -> Dict[str, float]:
+        metrics: Dict[str, float] = {}
+
+        if self.log_scope.get("global", True):
+            by_kind: Dict[str, Dict[str, Dict[str, float]]] = {}
+            for k, v in self.accumulator.tensors.items():
+                by_kind.setdefault(self._kind_of(k), {})[k] = v
+            for kind, entries in by_kind.items():
+                stats = self.accumulator._aggregate(entries)
+                for kk, vv in stats.items():
+                    metrics[f"{scope}/global/{kind}/{kk}"] = vv
+
+        if self.log_scope.get("per_layer", True):
+            layer_prefixes = set()
+            for name in self.accumulator.tensors:
+                clean = self._strip_kind(name)
+                parts = clean.split(".")
+                for i, p in enumerate(parts):
+                    if p == "layers" and i + 1 < len(parts):
+                        prefix = ".".join(parts[: i + 2])
+                        layer_prefixes.add(prefix)
+            for prefix in layer_prefixes:
+                by_kind: Dict[str, Dict[str, Dict[str, float]]] = {}
+                for k, v in self.accumulator.tensors.items():
+                    clean = self._strip_kind(k)
+                    if clean.startswith(prefix + ".") or clean == prefix:
+                        by_kind.setdefault(self._kind_of(k), {})[k] = v
+                safe = prefix.replace(".", "_")
+                for kind, entries in by_kind.items():
+                    if not entries:
+                        continue
+                    stats = self.accumulator._aggregate(entries)
+                    for kk, vv in stats.items():
+                        metrics[f"{scope}/layer_{safe}/{kind}/{kk}"] = vv
+
+        if self.log_scope.get("per_tensor", False):
+            for name, stats in self.accumulator.tensors.items():
+                safe = name.replace(".", "_")
+                for kk, vv in stats.items():
+                    if kk == "numel":
+                        continue
+                    metrics[f"{scope}/tensor_{safe}/{kk}"] = vv
+
+        return metrics
+
+    def on_log(self, args, state, control, logs=None, **kwargs):
+        if logs is not None and self.pending_metrics is not None:
+            logs.update(self.pending_metrics)
+            self.pending_metrics = None
+
+    def on_prediction_step(self, args, state, control, **kwargs):
+        if not self.monitor_during_eval:
+            return
+        if not self.hooks.active:
+            self.accumulator.clear()
+            self.hooks.set_active(True)
+            for name, param in self.model.named_parameters():
+                stats = StatsEngine.compute(
+                    param.data, self.user_limits["param"], self.dtype_ratio
+                )
+                self.accumulator.add(f"param.{name}", param.numel(), stats)
+        self.pending_metrics = self._build_metrics(scope="eval")
+
+    def on_evaluate(self, args, state, control, metrics=None, **kwargs):
+        self.hooks.set_active(False)
+        self.accumulator.clear()
+
+
+class TimeTrackerCallback(TrainerCallback):
+    """Precise training & eval timing with remaining-time estimates."""
+
+    def __init__(self):
+        self.step_start: Optional[float] = None
+        self.epoch_start: Optional[float] = None
+        self.total_train_time = 0.0
+        self.step_times: List[float] = []
+
+    def on_epoch_begin(self, args, state, control, **kwargs):
+        self.epoch_start = time.perf_counter()
+
+    def on_step_begin(self, args, state, control, **kwargs):
+        self.step_start = time.perf_counter()
+
+    def on_step_end(self, args, state, control, **kwargs):
+        if self.step_start is not None:
+            dt = time.perf_counter() - self.step_start
+            self.step_times.append(dt)
+            self.total_train_time += dt
+            self.step_start = None
+
+    def on_log(self, args, state, control, logs=None, **kwargs):
+        if logs is None:
+            return
+        logs["train/total_time_seconds"] = self.total_train_time
+        if self.step_times:
+            recent = self.step_times[-100:]
+            logs["train/time_per_step_avg"] = sum(recent) / len(recent)
+        if self.epoch_start is not None:
+            logs["train/epoch_time_elapsed"] = time.perf_counter() - self.epoch_start
+        if state.max_steps and state.global_step > 0:
+            avg = self.total_train_time / state.global_step
+            remaining = (state.max_steps - state.global_step) * avg
+            logs["train/estimated_remaining_minutes"] = remaining / 60.0
+
+    def on_evaluate(self, args, state, control, metrics=None, **kwargs):
+        pass
+
+
+class MetricsLoggerCallback(TrainerCallback):
+    """Persist every logged dict as JSONL in the output dir."""
+
+    def __init__(self, output_dir: str):
+        self.output_dir = Path(output_dir)
+        self.output_dir.mkdir(parents=True, exist_ok=True)
+        self.log_file = self.output_dir / "training_log.jsonl"
+
+    def on_log(self, args, state, control, logs=None, **kwargs):
+        if logs is None:
+            return
+        entry = {
+            "step": state.global_step,
+            "epoch": state.epoch,
+            "timestamp": time.time(),
+            **logs,
+        }
+        with open(self.log_file, "a") as f:
+            f.write(json.dumps(entry, default=str) + "\n")
+
+
+# =============================================================================
+# CONTAMINATION CALLBACK
+# =============================================================================
+
+class ContaminationCallback(TrainerCallback):
+    """
+    Intentionally corrupt input_ids and labels for a window of steps.
+    Modes: "shift" (+1 mod vocab), "random" (uniform random IDs).
+    """
+    def __init__(
+        self,
+        vocab_size: int,
+        enabled: bool = False,
+        start_step: int = 0,
+        duration_steps: int = 0,
+        mode: str = "shift",
+        fraction: float = 1.0,
+        seed: Optional[int] = None,
+    ):
+        self.vocab_size = vocab_size
+        self.enabled = enabled
+        self.start_step = start_step
+        self.duration_steps = duration_steps
+        self.mode = mode
+        self.fraction = fraction
+        self.seed = seed
+        self.generator = torch.Generator()
+        if seed is not None:
+            self.generator.manual_seed(seed)
+        self._active = False
+
+    def _should_corrupt(self, state) -> bool:
+        if not self.enabled:
+            return False
+        step = state.global_step
+        return self.start_step <= step < self.start_step + self.duration_steps
+
+    def on_step_begin(self, args, state, control, **kwargs):
+        if not self._should_corrupt(state):
+            return
+        batch = kwargs.get("inputs")
+        if batch is None or not isinstance(batch, dict):
+            return
+        input_ids = batch.get("input_ids")
+        labels = batch.get("labels")
+        attention_mask = batch.get("attention_mask")
+        if input_ids is None or labels is None:
+            return
+
+        device = input_ids.device
+        batch_size, seq_len = input_ids.shape
+
+        if self.fraction >= 1.0:
+            corrupt_mask = torch.ones((batch_size, seq_len), dtype=torch.bool, device=device)
+        elif self.fraction <= 0.0:
+            return
+        else:
+            rand = torch.rand((batch_size, seq_len), generator=self.generator, device=device)
+            corrupt_mask = rand < self.fraction
+
+        if attention_mask is not None:
+            padding_mask = attention_mask == 0
+            corrupt_mask = corrupt_mask & (~padding_mask)
+
+        if not corrupt_mask.any():
+            return
+
+        if self.mode == "shift":
+            input_ids_corrupt = (input_ids + 1) % self.vocab_size
+            labels_corrupt = (labels + 1) % self.vocab_size
+        elif self.mode == "random":
+            random_ids = torch.randint(
+                0, self.vocab_size, input_ids.shape,
+                generator=self.generator, device=device
+            )
+            input_ids_corrupt = random_ids
+            labels_corrupt = random_ids
+        else:
+            raise ValueError(f"Unknown contamination mode: {self.mode}")
+
+        input_ids.masked_scatter_(corrupt_mask, input_ids_corrupt[corrupt_mask])
+        labels.masked_scatter_(corrupt_mask, labels_corrupt[corrupt_mask])
+        batch["input_ids"] = input_ids
+        batch["labels"] = labels
+
+        if not self._active:
+            print(f"[Contamination] Started at step {state.global_step} for {self.duration_steps} steps (mode={self.mode})")
+            self._active = True
+
+    def on_step_end(self, args, state, control, **kwargs):
+        if self._active and state.global_step >= self.start_step + self.duration_steps:
+            print(f"[Contamination] Ended at step {state.global_step}")
+            self._active = False
+
+
+# =============================================================================
+# 5. DATA & TRAINER FACTORY
+# =============================================================================
+
+def build_dataset(
+    tokenizer,
+    max_seq_len: int = 512,
+    split: str = "train",
+    dataset_name: str = "roneneldan/TinyStories",
+    max_samples: Optional[int] = None,
+):
+    """Concatenate and chunk TinyStories for causal LM."""
+    ds = load_dataset(dataset_name, split=split)
+    if max_samples is not None and split == "train":
+        ds = ds.select(range(min(max_samples, len(ds))))
+        print(f"[Dataset] Using first {len(ds)} samples for training (max_samples={max_samples})")
+
+    def tokenize(examples):
+        out = tokenizer(examples["text"], add_special_tokens=False)
+        eos_id = tokenizer.eos_token_id
+        out["input_ids"] = [ids + [eos_id] for ids in out["input_ids"]]
+        if "attention_mask" in out:
+            out["attention_mask"] = [mask + [1] for mask in out["attention_mask"]]
+        return out
+
+    tokenized = ds.map(
+        tokenize,
+        batched=True,
+        num_proc=4,
+        remove_columns=ds.column_names,
+        desc=f"Tokenizing {split}",
+    )
+
+    def group_texts(examples):
+        concatenated = {
+            k: list(chain.from_iterable(examples[k])) for k in examples.keys()
+        }
+        total_length = len(concatenated[list(examples.keys())[0]])
+        total_length = (total_length // max_seq_len) * max_seq_len
+        result = {
+            k: [t[i : i + max_seq_len] for i in range(0, total_length, max_seq_len)]
+            for k, t in concatenated.items()
+        }
+        result["labels"] = result["input_ids"].copy()
+        return result
+
+    return tokenized.map(
+        group_texts,
+        batched=True,
+        batch_size=10000,
+        num_proc=4,
+        desc=f"Chunking {split}",
+    )
+
+
+def create_trainer(
+    model,
+    tokenizer,
+    config: Dict[str, Any],
+    train_dataset,
+    eval_dataset=None,
+):
+    """Assemble HF Trainer with all custom callbacks."""
+    tc = config.get("training", {})
+    mc = config.get("monitor", {})
+
+    run_name = tc.get("run_name", None)
+
+    args = TrainingArguments(
+        output_dir=tc.get("output_dir", "./out"),
+        run_name=run_name,
+        num_train_epochs=tc.get("num_train_epochs", 3),
+        per_device_train_batch_size=tc.get("per_device_train_batch_size", 16),
+        per_device_eval_batch_size=tc.get("per_device_eval_batch_size", 16),
+        gradient_accumulation_steps=tc.get("gradient_accumulation_steps", 4),
+        learning_rate=tc.get("learning_rate", 3e-4),
+        weight_decay=tc.get("weight_decay", 0.0),
+        max_grad_norm=tc.get("max_grad_norm", 1.0),
+        optim=tc.get("optim", "adamw_torch"),
+        warmup_steps=tc.get("warmup_steps", 0),
+        lr_scheduler_type=tc.get("lr_scheduler_type", "cosine"),
+        bf16=tc.get("bf16", True),
+        logging_steps=tc.get("logging_steps", 10),
+        eval_strategy=tc.get("eval_strategy", "steps"),
+        eval_steps=tc.get("eval_steps", 500),
+        save_strategy=tc.get("save_strategy", "steps"),
+        save_steps=tc.get("save_steps", 1000),
+        load_best_model_at_end=tc.get("load_best_model_at_end", False),
+        report_to=tc.get("report_to", "tensorboard"),
+        push_to_hub=tc.get("push_to_hub", False),
+        hub_model_id=tc.get("hub_model_id", None),
+        hub_token=tc.get("hub_token") or os.environ.get("HF_TOKEN"),
+        max_steps=tc.get("max_steps", -1),
+        seed=tc.get("seed", 42),
+        data_seed=tc.get("data_seed", 42),
+        remove_unused_columns=False,
+    )
+
+    callbacks = [TimeTrackerCallback()]
+
+    cc = config.get("contamination", {})
+    if cc.get("enabled", False):
+        vocab_size = model.config.vocab_size
+        callbacks.append(
+            ContaminationCallback(
+                vocab_size=vocab_size,
+                enabled=True,
+                start_step=cc.get("start_step", 0),
+                duration_steps=cc.get("duration_steps", 0),
+                mode=cc.get("mode", "shift"),
+                fraction=cc.get("fraction", 1.0),
+                seed=cc.get("seed", None),
+            )
+        )
+
+    if mc.get("enabled", True):
+        callbacks.append(
+            StabilityMonitorCallback(
+                model=model,
+                monitor_every_n_steps=mc.get("monitor_every_n_steps", 10),
+                module_patterns=mc.get("module_patterns", [".*mlp.*", ".*self_attn.*", ".*residual.*"]),
+                user_limits=mc.get("user_limits", {"grad": 1.0, "param": 100.0, "act": 50.0}),
+                dtype_proximity_ratio=mc.get("dtype_proximity_ratio", 0.9),
+                log_scope=mc.get("log_scope", {"global": True, "per_layer": True, "per_tensor": False}),
+                monitor_during_eval=mc.get("monitor_during_eval", False),
+            )
+        )
+
+    callbacks.append(MetricsLoggerCallback(args.output_dir))
+
+    collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
+
+    trainer = Trainer(
+        model=model,
+        args=args,
+        train_dataset=train_dataset,
+        eval_dataset=eval_dataset,
+        data_collator=collator,
+        callbacks=callbacks,
+    )
+
+    try:
+        from transformers.integrations import get_reporting_integration_callbacks
+        reporting_types = tuple(get_reporting_integration_callbacks(args.report_to))
+    except Exception:
+        reporting_types = ()
+
+    if reporting_types:
+        handler = trainer.callback_handler
+        reporting_cbs = [cb for cb in handler.callbacks if isinstance(cb, reporting_types)]
+        other_cbs = [cb for cb in handler.callbacks if not isinstance(cb, reporting_types)]
+        handler.callbacks = other_cbs + reporting_cbs
+
+    return trainer

zain/Activation/llm_analyzer_wandb.py

@@ -0,0 +1,570 @@
+"""
+Multi-LLM Activation & Loss Analyzer with WandB Logging
+
+Evaluates multiple language models on WikiText dataset,
+computing per-tensor and global activation statistics (mean, max_abs, std, norm)
+and logging to Weights & Biases with separate runs per model.
+
+Requirements:
+    pip install transformers torch datasets wandb tqdm
+
+Usage:
+    export WANDB_PROJECT="llm-activation-analysis"
+    export WANDB_API_KEY="your-key"
+    python llm_analyzer_wandb.py
+"""
+
+import math
+import torch
+import torch.nn.functional as F
+from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig
+from datasets import load_dataset
+from typing import List, Dict, Optional, Union, Tuple
+from dataclasses import dataclass, asdict
+from collections import defaultdict
+import json
+import warnings
+import os
+from tqdm import tqdm
+
+import wandb
+
+warnings.filterwarnings("ignore")
+
+
+@dataclass
+class TensorStats:
+    mean: float
+    max_abs: float
+    std: float
+    norm: float
+    numel: int
+
+
+@dataclass
+class ModelResult:
+    model_name: str
+    loss: float
+    perplexity: float
+    global_act: TensorStats
+    layer_acts: Dict[str, TensorStats]
+    num_tokens: int
+    num_layers: int
+    hidden_size: int
+    num_params: int
+
+
+class ActivationHookManager:
+    """Manages forward hooks to capture activations from every tensor."""
+
+    def __init__(self):
+        self.activations = {}
+        self.hooks = []
+        # Set once per batch via set_attention_mask(); used to exclude
+        # padding-token positions from activation statistics.
+        self._attention_mask: Optional[torch.Tensor] = None
+
+    def set_attention_mask(self, attention_mask: Optional[torch.Tensor]):
+        """Call once per batch before the forward pass so hooks can mask
+        out padding positions when computing stats."""
+        self._attention_mask = (
+            attention_mask.detach().cpu() if attention_mask is not None else None
+        )
+
+    def _make_hook(self, name: str):
+        def hook(module, input, output):
+            # Handle different output types
+            if isinstance(output, torch.Tensor):
+                tensor = output
+            elif isinstance(output, tuple) and isinstance(output[0], torch.Tensor):
+                tensor = output[0]
+            else:
+                return
+
+            # Detach and move to CPU to avoid GPU memory blowup
+            self.activations[name] = tensor.detach().cpu().float()
+        return hook
+
+    def register_hooks(self, model: torch.nn.Module):
+        """Register hooks on all modules that produce activations."""
+        for name, module in model.named_modules():
+            # Skip trivial containers
+            if len(list(module.children())) == 0 and hasattr(module, 'forward'):
+                hook = module.register_forward_hook(self._make_hook(name))
+                self.hooks.append(hook)
+
+    def clear(self):
+        self.activations.clear()
+
+    def remove_hooks(self):
+        for hook in self.hooks:
+            hook.remove()
+        self.hooks.clear()
+
+    def _select_real_tokens(self, tensor: torch.Tensor) -> torch.Tensor:
+        """
+        BUG FIX: previously every captured tensor (including activations at
+        padding-token positions) was flattened and used as-is. With
+        padding_side="left" and a small batch_size, the padding fraction
+        varies a lot batch-to-batch, so padding-token activations (which
+        are real, non-zero values — not zeros) were silently mixed into
+        mean/std/max_abs/norm, biasing exactly the saturation signal this
+        script exists to measure.
+
+        Here we mask out padding positions whenever a tensor's shape is
+        consistent with (batch, seq_len, ...) against the stored
+        attention_mask (batch, seq_len). Tensors that don't match that
+        shape (e.g. a module operating on the pooled/final dimension only)
+        are left as-is rather than guessing.
+        """
+        mask = self._attention_mask
+        if mask is None or tensor.dim() < 2:
+            return tensor.reshape(-1)
+        if tensor.shape[0] != mask.shape[0] or tensor.shape[1] != mask.shape[1]:
+            return tensor.reshape(-1)
+
+        bool_mask = mask.bool()
+        # Expand mask across any trailing dims (e.g. hidden_size) and select.
+        expand_shape = bool_mask.shape + (1,) * (tensor.dim() - 2)
+        bool_mask = bool_mask.view(expand_shape).expand_as(tensor)
+        return tensor[bool_mask].reshape(-1)
+
+    def compute_stats(self) -> Dict[str, TensorStats]:
+        """Compute statistics for all captured activations, excluding
+        padding-token positions where identifiable."""
+        stats = {}
+        for name, tensor in self.activations.items():
+            if tensor.numel() == 0:
+                continue
+            flat = self._select_real_tokens(tensor)
+            if flat.numel() == 0:
+                continue
+            stats[name] = TensorStats(
+                mean=flat.mean().item(),
+                max_abs=flat.abs().max().item(),
+                std=flat.std().item(),
+                norm=flat.norm().item(),
+                numel=flat.numel()
+            )
+        return stats
+
+
+class LLMAnalyzer:
+    def __init__(
+        self,
+        device: Optional[str] = None,
+        max_length: int = 512,
+        max_samples: int = 1000,  # number of wikitext samples to eval
+        batch_size: int = 4,
+        dtype: torch.dtype = torch.float16,
+        wandb_project: Optional[str] = None,
+    ):
+        self.device = device or ("cuda" if torch.cuda.is_available() else "cpu")
+        self.max_length = max_length
+        self.max_samples = max_samples
+        self.batch_size = batch_size
+        self.dtype = dtype if self.device == "cuda" else torch.float32
+        self.wandb_project = wandb_project or os.environ.get("WANDB_PROJECT", "llm-activation-analysis")
+        self._cache = {}
+
+    def load_dataset(self, split: str = "test"):
+        """Load Salesforce/wikitext dataset."""
+        print(f"[Dataset] Loading Salesforce/wikitext ({split}) ...")
+        ds = load_dataset("Salesforce/wikitext", "wikitext-2-raw-v1", split=split)
+        # Filter out empty lines
+        texts = [t for t in ds["text"] if len(t.strip()) > 50]
+        print(f"[Dataset] Loaded {len(texts)} non-empty samples")
+        return texts[:self.max_samples]
+
+    def load_model(self, model_name: str):
+        """Load model and tokenizer with caching."""
+        if model_name in self._cache:
+            return self._cache[model_name]
+
+        print(f"[Loading] {model_name} ...")
+
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_name,
+            trust_remote_code=True,
+            padding_side="left"
+        )
+        if tokenizer.pad_token is None:
+            tokenizer.pad_token = tokenizer.eos_token
+
+        config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
+
+        model = AutoModelForCausalLM.from_pretrained(
+            model_name,
+            config=config,
+            torch_dtype=self.dtype,
+            device_map="auto" if self.device == "cuda" else None,
+            trust_remote_code=True,
+        )
+
+        if self.device == "cpu":
+            model = model.to(self.device)
+
+        model.eval()
+
+        num_params = sum(p.numel() for p in model.parameters())
+
+        self._cache[model_name] = (tokenizer, model, config, num_params)
+        print(f"[Loaded]  {model_name} | Params: {num_params/1e6:.1f}M | Layers: {config.num_hidden_layers} | Hidden: {config.hidden_size}")
+        return tokenizer, model, config, num_params
+
+    def compute(
+        self,
+        model_names: List[str],
+    ) -> List[ModelResult]:
+        """
+        Compute loss and per-tensor activation statistics for multiple models.
+        Logs each model as a separate WandB run.
+        """
+        texts = self.load_dataset()
+        results = []
+
+        for model_name in model_names:
+            try:
+                result = self._evaluate_model(model_name, texts)
+                results.append(result)
+            except Exception as e:
+                print(f"[Error] {model_name}: {e}")
+                import traceback
+                traceback.print_exc()
+                continue
+
+        return results
+
+    def _evaluate_model(
+        self,
+        model_name: str,
+        texts: List[str],
+    ) -> ModelResult:
+        tokenizer, model, config, num_params = self.load_model(model_name)
+
+        # Initialize WandB run for this model
+        run_name = model_name.replace("/", "-")
+        wandb.init(
+            project=self.wandb_project,
+            name=run_name,
+            config={
+                "model": model_name,
+                "max_length": self.max_length,
+                "max_samples": self.max_samples,
+                "batch_size": self.batch_size,
+                "dtype": str(self.dtype),
+                "num_params": num_params,
+                "num_layers": config.num_hidden_layers,
+                "hidden_size": config.hidden_size,
+            },
+            reinit=True
+        )
+
+        hook_mgr = ActivationHookManager()
+        hook_mgr.register_hooks(model)
+
+        total_loss = 0.0
+        total_tokens = 0
+
+        # Global activation accumulator
+        global_acts = []
+
+        # Per-layer activation accumulators
+        # We'll aggregate stats across batches, then compute final stats
+        layer_act_values = defaultdict(list)
+
+        num_batches = (len(texts) + self.batch_size - 1) // self.batch_size
+
+        for i in tqdm(range(0, len(texts), self.batch_size), desc=f"Eval {run_name}", total=num_batches):
+            batch_texts = texts[i:i + self.batch_size]
+
+            # Tokenize
+            inputs = tokenizer(
+                batch_texts,
+                return_tensors="pt",
+                truncation=True,
+                max_length=self.max_length,
+                padding=True
+            )
+
+            # Move to device
+            if self.device == "cuda" and hasattr(model, "device"):
+                # model is on auto device map
+                input_ids = inputs["input_ids"]
+                if hasattr(model, "device") and model.device != torch.device("meta"):
+                    input_ids = input_ids.to(model.device)
+                attention_mask = inputs.get("attention_mask")
+                if attention_mask is not None:
+                    attention_mask = attention_mask.to(input_ids.device)
+            else:
+                input_ids = inputs["input_ids"].to(self.device)
+                attention_mask = inputs.get("attention_mask")
+                if attention_mask is not None:
+                    attention_mask = attention_mask.to(self.device)
+
+            labels = input_ids.clone()
+
+            # BUG FIX: with padding_side="left" and no explicit position_ids,
+            # HF's default `position_ids = arange(seq_len)` is applied
+            # identically to every row in the batch, regardless of how much
+            # left-padding precedes the real tokens in that row (verified
+            # against transformers' LlamaModel.forward / GPT2Model.forward
+            # source — neither adjusts for padding when position_ids=None).
+            # That means a real token's absolute position (and therefore its
+            # RoPE rotation / absolute position embedding) depends on how
+            # much padding happened to precede it in this particular batch,
+            # not on its logical position within its own sequence. This
+            # silently corrupts logits -> loss -> perplexity, with the
+            # amount of corruption varying batch-to-batch. Fix: derive
+            # position_ids from attention_mask so they restart at 0 for the
+            # first real token of every row, and are stable (0) on padding.
+            if attention_mask is not None:
+                position_ids = attention_mask.long().cumsum(-1) - 1
+                position_ids.masked_fill_(attention_mask == 0, 0)
+            else:
+                position_ids = None
+
+            hook_mgr.set_attention_mask(attention_mask)
+
+            with torch.no_grad():
+                outputs = model(
+                    input_ids=input_ids,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    labels=labels,
+                )
+
+            # --- Loss Computation ---
+            logits = outputs.logits
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            shift_mask = attention_mask[..., 1:].contiguous() if attention_mask is not None else None
+
+            loss_fct = torch.nn.CrossEntropyLoss(reduction="none")
+            token_losses = loss_fct(
+                shift_logits.view(-1, shift_logits.size(-1)),
+                shift_labels.view(-1)
+            )
+
+            if shift_mask is not None:
+                token_losses = token_losses * shift_mask.view(-1)
+                num_valid_tokens = shift_mask.sum().item()
+            else:
+                num_valid_tokens = token_losses.numel()
+
+            batch_loss = token_losses.sum().item()
+            total_loss += batch_loss
+            total_tokens += num_valid_tokens
+
+            # --- Activation Statistics ---
+            # Get activations captured by hooks
+            act_stats = hook_mgr.compute_stats()
+
+            for name, stats in act_stats.items():
+                # Skip dtype-limit fraction metrics entirely
+                # We only log mean, max_abs, std, norm
+
+                # For global: aggregate raw values
+                # We can't store all raw values due to memory, so we store running sums
+                # But for accurate std across all batches, we need a streaming algorithm
+                # For simplicity and correctness, we'll store per-batch stats and weight them
+                layer_act_values[name].append(asdict(stats))
+
+            hook_mgr.clear()
+
+            # Log per-batch metrics to wandb
+            if num_valid_tokens > 0:
+                batch_avg_loss = batch_loss / num_valid_tokens
+                wandb.log({
+                    "batch_loss": batch_avg_loss,
+                    "batch_perplexity": torch.exp(torch.tensor(batch_avg_loss)).item(),
+                    "batch_tokens": num_valid_tokens,
+                    "progress": i / len(texts)
+                }, step=i)
+
+        hook_mgr.remove_hooks()
+
+        # --- Final Aggregation ---
+        avg_loss = total_loss / max(total_tokens, 1)
+        perplexity = torch.exp(torch.tensor(avg_loss)).item()
+
+        # Aggregate per-tensor stats across all batches
+        # Weighted by numel for mean, max for max_abs, pooled std, pooled norm
+        final_layer_stats = {}
+
+        for name, batch_stats_list in layer_act_values.items():
+            total_numel = sum(s["numel"] for s in batch_stats_list)
+            if total_numel == 0:
+                continue
+
+            # Weighted mean
+            weighted_mean = sum(s["mean"] * s["numel"] for s in batch_stats_list) / total_numel
+
+            # Max abs across all batches
+            max_abs = max(s["max_abs"] for s in batch_stats_list)
+
+            # BUG FIX: the previous formula (weighted average of per-batch
+            # variances only) drops the between-batch term that accounts for
+            # per-batch means differing from the global mean. Whenever batch
+            # means differ (they will — different texts, different lengths),
+            # this systematically UNDERESTIMATES the true global std — in a
+            # quick numeric test with two batches of different means this was
+            # off by ~2.8x. Correct pooled-variance formula (population form,
+            # matches exp.py's StepAccumulator._merge_entry):
+            #   E[X^2] = weighted_avg(var_i + mean_i^2)
+            #   Var(X) = E[X^2] - mean_global^2
+            ex2 = sum(
+                s["numel"] * (s["std"] ** 2 + s["mean"] ** 2) for s in batch_stats_list
+            ) / total_numel
+            pooled_std = math.sqrt(max(0.0, ex2 - weighted_mean ** 2))
+
+            # Norm: sqrt(sum of squared norms / total_numel) * sqrt(total_numel)
+            # Actually norm^2 = sum(x_i^2), so pooled_norm = sqrt(sum(norm_i^2))
+            pooled_norm = (sum(s["norm"] ** 2 for s in batch_stats_list)) ** 0.5
+
+            final_layer_stats[name] = TensorStats(
+                mean=weighted_mean,
+                max_abs=max_abs,
+                std=pooled_std,
+                norm=pooled_norm,
+                numel=total_numel
+            )
+
+        # Compute global stats across all layers
+        if final_layer_stats:
+            all_numel = sum(s.numel for s in final_layer_stats.values())
+            global_mean = sum(s.mean * s.numel for s in final_layer_stats.values()) / all_numel
+            global_max_abs = max(s.max_abs for s in final_layer_stats.values())
+            # Same pooled-std correction as above — same bug was present here.
+            global_ex2 = sum(
+                s.numel * (s.std ** 2 + s.mean ** 2) for s in final_layer_stats.values()
+            ) / all_numel
+            global_std = math.sqrt(max(0.0, global_ex2 - global_mean ** 2))
+            global_norm = (sum(s.norm ** 2 for s in final_layer_stats.values())) ** 0.5
+
+            global_stats = TensorStats(
+                mean=global_mean,
+                max_abs=global_max_abs,
+                std=global_std,
+                norm=global_norm,
+                numel=all_numel
+            )
+        else:
+            global_stats = TensorStats(0.0, 0.0, 0.0, 0.0, 0)
+
+        result = ModelResult(
+            model_name=model_name,
+            loss=avg_loss,
+            perplexity=perplexity,
+            global_act=global_stats,
+            layer_acts=final_layer_stats,
+            num_tokens=total_tokens,
+            num_layers=config.num_hidden_layers,
+            hidden_size=config.hidden_size,
+            num_params=num_params
+        )
+
+        # --- WandB Logging ---
+        self._log_to_wandb(result)
+        wandb.finish()
+
+        return result
+
+    def _log_to_wandb(self, result: ModelResult):
+        """Log final metrics to WandB. No frac_near_dtype_limit."""
+
+        # Global metrics
+        wandb.log({
+            "final/loss": result.loss,
+            "final/perplexity": result.perplexity,
+            "final/num_tokens": result.num_tokens,
+
+            "train/global/act/mean": result.global_act.mean,
+            "train/global/act/max_abs": result.global_act.max_abs,
+            "train/global/act/std": result.global_act.std,
+            "train/global/act/norm": result.global_act.norm,
+            # Intentionally NOT logging frac_near_dtype_limit or frac_near_user_limit
+        })
+
+        # Per-tensor (per-layer) metrics
+        # Organize by layer for cleaner WandB UI
+        for tensor_name, stats in result.layer_acts.items():
+            # Clean name for wandb: replace dots with slashes
+            clean_name = tensor_name.replace(".", "/")
+
+            wandb.log({
+                f"train/{clean_name}/act/mean": stats.mean,
+                f"train/{clean_name}/act/max_abs": stats.max_abs,
+                f"train/{clean_name}/act/std": stats.std,
+                f"train/{clean_name}/act/norm": stats.norm,
+                # No frac_near_dtype_limit
+            })
+
+        # Also log as a wandb.Table for easy comparison
+        table_data = []
+        for tensor_name, stats in sorted(result.layer_acts.items()):
+            table_data.append([
+                tensor_name,
+                stats.mean,
+                stats.max_abs,
+                stats.std,
+                stats.norm,
+                stats.numel
+            ])
+
+        if table_data:
+            table = wandb.Table(
+                columns=["tensor_name", "mean", "max_abs", "std", "norm", "numel"],
+                data=table_data
+            )
+            wandb.log({"activation_table": table})
+
+    def print_report(self, results: List[ModelResult]):
+        """Pretty-print comparison report."""
+        print("\n" + "=" * 110)
+        print(f"{'Model':<35} {'Loss':>10} {'PPL':>10} {'ActMean':>12} {'ActMaxAbs':>12} {'ActStd':>12} {'Tokens':>8}")
+        print("-" * 110)
+
+        for r in results:
+            name = r.model_name.split("/")[-1][:33]
+            print(
+                f"{name:<35} "
+                f"{r.loss:>10.4f} "
+                f"{r.perplexity:>10.2f} "
+                f"{r.global_act.mean:>12.6f} "
+                f"{r.global_act.max_abs:>12.6f} "
+                f"{r.global_act.std:>12.6f} "
+                f"{r.num_tokens:>8}"
+            )
+
+        print("=" * 110)
+
+        # Print top 5 layers by max_abs for each model
+        print("\n[Per-Tensor Max Abs Top 5]")
+        for r in results:
+            name = r.model_name.split("/")[-1]
+            sorted_layers = sorted(r.layer_acts.items(), key=lambda x: x[1].max_abs, reverse=True)[:5]
+            print(f"\n  {name}:")
+            for tensor_name, stats in sorted_layers:
+                print(f"    {tensor_name:<50} max_abs={stats.max_abs:>10.4f}  mean={stats.mean:>10.6f}  std={stats.std:>10.4f}")
+
+    def export_json(self, results: List[ModelResult], path: str):
+        """Export results to JSON."""
+        data = []
+        for r in results:
+            entry = {
+                "model": r.model_name,
+                "loss": r.loss,
+                "perplexity": r.perplexity,
+                "num_tokens": r.num_tokens,
+                "num_layers": r.num_layers,
+                "hidden_size": r.hidden_size,
+                "num_params": r.num_params,
+                "global_act": asdict(r.global_act),
+                "layer_acts": {k: asdict(v) for k, v in r.layer_acts.items()}
+            }
+            data.append(entry)
+
+        with open(path, "w") as f:
+            json.dump(data, f, indent=2)
+        print(f"[Exported] Results saved to {path}")

zain/Activation/sweep.py

@@ -0,0 +1,160 @@
+#!/usr/bin/env python3
+"""Sweep explicit GLU / MLP variants with identical data and hyperparameters."""
+import argparse
+import copy
+import json
+import re
+import time
+import os
+from pathlib import Path
+
+import yaml
+import wandb
+import torch
+from transformers import AutoTokenizer, set_seed
+from exp import TinyLlamaConfig, TinyLlamaForCausalLM, build_dataset, create_trainer
+
+
+def format_param_count(total_params: int) -> str:
+    """Return human‑readable string with M or B suffix, 1 decimal."""
+    if total_params >= 1e9:
+        return f"{total_params / 1e9:.1f}B"
+    else:
+        return f"{total_params / 1e6:.1f}M"
+
+
+def parse_variant(variant: str):
+    """
+    Parse variant string into (prefix, activation, layers).
+    Supports:
+        glu-silu-waleed10
+        glu-situglu_low-10L
+        mlp-waleed10
+    """
+    parts = variant.split('-')
+    if len(parts) < 2:
+        raise ValueError(f"Invalid variant format: '{variant}'. Expected: <glu|mlp>-<activation>[-<layers>L]")
+
+    prefix = parts[0]
+    if prefix not in ('glu', 'mlp'):
+        raise ValueError(f"Invalid prefix: '{prefix}'. Must be 'glu' or 'mlp'.")
+
+    last = parts[-1]
+    if last.endswith('L') and last[:-1].isdigit():
+        layers = int(last[:-1])
+        activation = '-'.join(parts[1:-1])
+    else:
+        layers = None
+        activation = '-'.join(parts[1:])
+
+    if not activation:
+        raise ValueError(f"Missing activation name in variant: '{variant}'")
+
+    return prefix, activation, layers
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--config", required=True, help="Base YAML config")
+    parser.add_argument(
+        "--variants",
+        nargs="+",
+        required=True,
+        help="List of variants: e.g. glu-silu-10L mlp-relu-8L"
+    )
+    parser.add_argument("--push", action="store_true")
+    args = parser.parse_args()
+
+    with open(args.config) as f:
+        base = yaml.safe_load(f)
+
+    seed = base.get("training", {}).get("seed", 42)
+    set_seed(seed)
+
+    # Force all runs into a specific WandB project (read from config)
+    wandb_project = base.get("training", {}).get("wandb_project")
+    if wandb_project:
+        os.environ["WANDB_PROJECT"] = wandb_project
+        print(f"[WandB] Project locked to: {wandb_project}")
+
+    tok_name = base["model"].get("tokenizer_name", "meta-llama/Llama-2-7b-hf")
+    tokenizer = AutoTokenizer.from_pretrained(tok_name)
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
+
+    msl = base["model"].get("max_position_embeddings", 512)
+    train_ds = build_dataset(tokenizer, max_seq_len=msl, split="train", max_samples=None)
+    eval_ds = build_dataset(tokenizer, max_seq_len=msl, split="validation", max_samples=None)
+
+    results = []
+
+    for variant in args.variants:
+        prefix, act, layers = parse_variant(variant)
+
+        if prefix == "mlp" and act in ("situglu", "waleed", "situglu_low", "waleedglu_low"):
+            raise ValueError(
+                f"Activation '{act}' requires a gated architecture (GLU). "
+                f"Please use 'glu-{act}' instead."
+            )
+
+        cfg = copy.deepcopy(base)
+        cfg["model"]["mlp_type"] = prefix
+        cfg["model"]["activation"] = act
+        if layers is not None:
+            cfg["model"]["num_hidden_layers"] = layers
+
+        # --------------------------------------------------------------
+        # ENSURE LAYER COUNT IS ALWAYS INCLUDED IN ALL NAMES
+        # Read the final layer count from the config after any override.
+        # --------------------------------------------------------------
+        actual_layers = cfg["model"]["num_hidden_layers"]
+        variant_label = f"{prefix}-{act}-{actual_layers}L"
+
+        out_dir = Path(cfg["training"]["output_dir"]).parent / f"{variant_label}_run"
+        cfg["training"]["output_dir"] = str(out_dir)
+
+        set_seed(seed)
+
+        print(f"\n{'='*60}\n>>> Variant: {variant_label} | Out: {out_dir}\n{'='*60}")
+
+        config = TinyLlamaConfig(**cfg["model"])
+        model = TinyLlamaForCausalLM(config)
+        model = model.to(torch.bfloat16)
+
+        total_params = sum(p.numel() for p in model.parameters())
+        param_str = format_param_count(total_params)
+        timestamp = time.strftime("%Y%m%d-%H%M%S")
+        run_name = f"LM-{variant_label}-{param_str}-{timestamp}"
+        cfg["training"]["run_name"] = run_name
+
+        hub_id_base = cfg["training"].get("hub_model_id", "tiny-llama-lab")
+        cfg["training"]["hub_model_id"] = f"{hub_id_base}-{variant_label}"
+
+        os.environ.pop("WANDB_RUN_ID", None)
+
+        trainer = create_trainer(model, tokenizer, cfg, train_ds, eval_ds)
+
+        try:
+            trainer.train()
+            metrics = trainer.evaluate()
+            results.append({
+                "variant": variant_label,
+                "eval_loss": metrics.get("eval_loss"),
+                "out": str(out_dir),
+                "run_name": run_name,
+            })
+            trainer.save_model(str(out_dir))
+            if args.push or cfg["training"].get("push_to_hub", False):
+                trainer.push_to_hub()
+        finally:
+            wandb.finish()
+
+    summary = Path(base["training"]["output_dir"]).parent / "sweep_summary.json"
+    summary.write_text(json.dumps(results, indent=2))
+    print("\nSweep complete:")
+    for r in results:
+        print(f"  {r['variant']:20s}  eval_loss={r['eval_loss']:.4f}")
+
+
+if __name__ == "__main__":
+    main()

zain/Activation/train.py

@@ -0,0 +1,63 @@
+#!/usr/bin/env python3
+"""Train one TinyLlama variant from a YAML config."""
+import argparse
+import yaml
+import os
+import torch
+
+from transformers import AutoTokenizer, set_seed
+from exp import TinyLlamaConfig, TinyLlamaForCausalLM, build_dataset, create_trainer
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--config", required=True, help="Path to YAML config")
+    parser.add_argument("--push", action="store_true", help="Push final model to HF Hub")
+    args = parser.parse_args()
+
+    with open(args.config) as f:
+        cfg = yaml.safe_load(f)
+
+    seed = cfg.get("training", {}).get("seed", 42)
+    set_seed(seed)
+
+    # -------------------------------------------------------------------------
+    # LOCK WANDB PROJECT: read from config, force into environment.
+    # -------------------------------------------------------------------------
+    wandb_project = cfg.get("training", {}).get("wandb_project")
+    if wandb_project:
+        os.environ["WANDB_PROJECT"] = wandb_project
+        print(f"[WandB] Project locked to: {wandb_project}")
+    # -------------------------------------------------------------------------
+
+    model_cfg = cfg["model"]
+    train_cfg = cfg.get("training", {})
+
+    tok_name = model_cfg.pop("tokenizer_name", "meta-llama/Llama-2-7b-hf")
+    tokenizer = AutoTokenizer.from_pretrained(tok_name)
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
+
+    tiny_config = TinyLlamaConfig(**model_cfg)
+    model = TinyLlamaForCausalLM(tiny_config)
+    model = model.to(torch.bfloat16)
+
+    n_params = sum(p.numel() for p in model.parameters()) / 1e6
+    print(f"Model: {n_params:.2f}M params | MLP type: {tiny_config.mlp_type} | Activation: {tiny_config.activation}")
+
+    msl = model_cfg.get("max_position_embeddings", 512)
+    train_ds = build_dataset(tokenizer, max_seq_len=msl, split="train", max_samples=None)
+    eval_ds = build_dataset(tokenizer, max_seq_len=msl, split="validation", max_samples=None)
+
+    trainer = create_trainer(model, tokenizer, cfg, train_ds, eval_ds)
+    trainer.train()
+
+    out = train_cfg.get("output_dir", "./out")
+    trainer.save_model(out)
+    if args.push or train_cfg.get("push_to_hub", False):
+        trainer.push_to_hub()
+    print(f"Done. Artifacts in {out}")
+
+
+if __name__ == "__main__":
+    main()