created micro hf space

.gitignore

@@ -0,0 +1,4 @@
+.env
+*.pyc
+.DS_Store
+__pycache__/

Instructions.md

@@ -0,0 +1,20 @@
+# MiCRo Expert Routing Visualizer (Gradio)
+
+This demo visualizes how modular language models allocate computation across specialized experts—Language, Logic, Social, and World—when processing a given prompt.
+Each expert corresponds to a cognitive domain inspired by human brain networks. Enter a prompt to see how tokens are dynamically routed across modules, revealing the model’s internal reasoning structure.
+
+## How it works
+- Choose a model (dropdown) or type a custom model id.
+- Enter a *User prompt*. Optionally add an *Assistant prompt*; if provided, the app concatenates them as:
+
+  ```
+  User: <user text>
+  Assistant: <assistant text>
+  ```
+
+- When the prompt fails, the demo falls back to "mock data", which generates deterministic, pseudo-random percentages from the prompt.  
+
+### Backend contract
+`get_expert_routing(model_id: str, prompt: str)` must return 4 values (percentages) for the experts in this fixed order:
+`["Language", "Logic", "Social", "World"]`
+or a dict with those exact keys.

README.md

@@ -1,6 +1,5 @@
----
 title: MiCRo Routing Visualizer
-emoji: 💻
+emoji: 🧠
 colorFrom: purple
 colorTo: red
 sdk: gradio
@@ -8,7 +7,4 @@ sdk_version: 5.49.1
 app_file: app.py
 pinned: false
 license: mit
-short_description: Mixture of Cognitive Reasoners Computation Allocation
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+short_description: Mixture of Cognitive Reasoners Computation Allocation

app.py

@@ -0,0 +1,166 @@
+# app.py
+"""
+Hugging Face Space: MoE Expert Routing Visualizer (Gradio)
+----------------------------------------------------------
+This Space lets a user:
+- Choose a model (from a dropdown or a free-text box)
+- Enter a user prompt, and optionally an assistant prompt
+- Call a backend function that returns 4 routing percentages (Language, Logic, Social, World)
+- See a bar plot + table of the percentages
+
+🧩 Plug your real routing function in router_backend.py -> get_expert_routing().
+By default, a deterministic "mock mode" produces stable pseudo-random percentages from the prompt.
+"""
+
+import hashlib
+from typing import Dict, List, Tuple, Union
+import gradio as gr
+import plotly
+import plotly.express as px
+import pandas as pd
+from router_backend import get_expert_routing 
+
+# ---- Expected backend adapter ------------------------------------------------
+# Implement your real function in router_backend.py with the following signature:
+#   def get_expert_routing(model_id: str, prompt: str) -> Union[List[float], Dict[str, float], Tuple[float, float, float, float]]
+# It MUST return 4 values that sum to ~100 (percentages) in the fixed order:
+#   ["Language", "Logic", "Social", "World"]
+# or a mapping with those keys.
+# try:
+#     from router_backend import get_expert_routing  # your real backend
+#     BACKEND_AVAILABLE = True
+# except Exception as e:  # keep error for display if needed
+#     BACKEND_AVAILABLE = False
+#     _backend_import_error = e
+
+EXPERTS = ["Language", "Logic", "Social", "World"]
+
+DEFAULT_MODELS = [
+    "micro-llama-1b", 
+    "micro-llama-3b",
+    "micro-llama-1b-dpo",
+    "micro-moe-llama-1b",
+    "micro-smollm2-135m",
+    "micro-smollm2-360m",
+    "micro-moe-smollm2-135m",
+    "micro-moe-smollm2-360m",
+]
+
+def _mock_routing(model_id: str, prompt: str, seed: int = 0) -> List[float]:
+    """
+    Deterministic mock routing percentages based on model_id + prompt + seed.
+    Returns a list of 4 percentages summing to 100.0
+    """
+    h = hashlib.sha256(f"{model_id}||{prompt}||{seed}".encode()).digest()
+    # split into 4 positive numbers
+    vals = [int.from_bytes(h[i*8:(i+1)*8], "little") % 10_000 + 1 for i in range(4)]
+    s = sum(vals)
+    return [100.0 * v / s for v in vals]
+
+def _normalize_output(r: Union[List[float], Tuple[float, float, float, float], Dict[str, float]]) -> List[float]:
+    """
+    Normalize different return types into a 4-length list ordered as EXPERTS.
+    """
+    if isinstance(r, dict):
+        vals = [float(r.get(k, 0.0)) for k in EXPERTS]
+    else:
+        vals = [float(x) for x in list(r)]
+        if len(vals) != 4:
+            raise ValueError(f"Expected 4 values, got {len(vals)}.")
+    # renormalize to 100 if needed
+    s = sum(vals)
+    if s <= 0:
+        raise ValueError("Sum of routing percentages is non-positive.")
+    vals = [100.0 * v / s for v in vals]
+    return vals
+
+def _compose_prompt(user_prompt: str, assistant_prompt: str) -> str:
+    user_prompt = (user_prompt or "").strip()
+    assistant_prompt = (assistant_prompt or "").strip()
+    if assistant_prompt:
+        return [{"role": "user", "content": user_prompt}, {"role": "assistant", "content": assistant_prompt}]
+    return user_prompt
+
+def route_and_plot(model_choice: str, hf_token: str, user_prompt: str, assistant_prompt: str) -> Tuple[pd.DataFrame, "plotly.graph_objs._figure.Figure", str]:
+    """
+    Main pipeline:
+    - Compose prompt (user + optional assistant)
+    - Call backend (real or mock)
+    - Return a table + bar plot + status message
+    """
+    model_id = model_choice.strip()
+    if not model_id:
+        raise gr.Error("Please select a model or enter a custom model id.")
+    prompt = _compose_prompt(user_prompt, assistant_prompt)
+    if not prompt:
+        raise gr.Error("Please enter a prompt.")
+    
+    seed = 42
+    use_mock = False
+    if use_mock:
+        msg = "Using mock data."
+        vals = _mock_routing(model_id, prompt, seed=seed)
+        generation = None
+    else:
+        try:
+            raw, generation = get_expert_routing(model_id, hf_token, prompt)  # <-- your real function
+            vals = _normalize_output(raw)
+            msg = "Routed with real backend."
+        except Exception as e:
+            # fallback to mock on error, but surface message
+            msg = f"Backend error: {e}\nFalling back to mock data."
+            vals = _mock_routing(model_id, prompt, seed=seed)
+            generation = None
+
+    df = pd.DataFrame({"Expert": EXPERTS, "Percent": vals})
+    fig = px.bar(df, x="Expert", y="Percent", title="Token Routing by Expert (%)", text="Percent")
+    fig.update_traces(texttemplate="%{text:.2f}%", textposition="outside")
+    fig.update_layout(yaxis_range=[0, max(100, max(vals) * 1.25)], bargap=0.35)
+
+    status = f"Model: {model_id}<br>{msg}"
+    if generation is None:
+        generation = assistant_prompt
+
+    return generation, df, fig, status
+
+with gr.Blocks(theme=gr.themes.Soft(), fill_height=True) as demo:
+    gr.Markdown(
+        """
+        # 🧠 Mixture of Cognitive Reasoner (MiCRo) Expert Routing Visualizer
+        ## Enter a prompt (and optionally an assistant reply), pick a model, and visualize how tokens were routed across experts.
+        Paper: [Mixture of Cognitive Reasoners: Modular Reasoning with Brain-Like Specialization](https://arxiv.org/abs/2506.13331)
+        ----
+        This demo visualizes how modular language models allocate computation across specialized experts—Language, Logic, Social, and World—when processing a given prompt.
+        Each expert corresponds to a cognitive domain inspired by human brain networks. Enter a prompt to see how tokens are dynamically routed across modules, revealing the model's internal reasoning structure.
+        """.strip()
+    )
+
+    with gr.Row():
+        model_choice = gr.Dropdown(choices=DEFAULT_MODELS, label="Select a model", value=DEFAULT_MODELS[0])
+        hf_token = gr.Textbox(label="Huggingface token for authentication", placeholder="hf token", lines=1)
+
+    with gr.Row():
+        user_prompt = gr.Textbox(lines=6, label="User prompt", placeholder="Type the user message here...")
+        assistant_prompt = gr.Textbox(lines=6, label="Assistant prompt (optional)", placeholder="Type the assistant message here (optional)...")
+
+    # with gr.Row():
+    #     use_mock = gr.Checkbox(value=True, label="Use mock data (uncheck to call your backend)")
+    #     seed = gr.Slider(value=0, minimum=0, maximum=10_000, step=1, label="Mock seed")
+
+    run = gr.Button("Run Routing", variant="primary")
+
+    generation_output = gr.Textbox(lines=4, label="Generated continuation", placeholder="Generated text will appear here...", interactive=False)
+
+    with gr.Row():
+        table = gr.Dataframe(label="Routing Percentages", interactive=False)
+    plot = gr.Plot(label="Bar Plot")
+    status = gr.Markdown("")
+
+    run.click(
+        route_and_plot,
+        inputs=[model_choice, hf_token, user_prompt, assistant_prompt],
+        outputs=[generation_output, table, plot, status],
+    )
+
+if __name__ == "__main__":
+    demo.launch()

configs/micro_llama_1b.yml

@@ -0,0 +1,14 @@
+run-title: micro-llama-1b
+model: micro-llama-1b
+
+base-model: meta-llama/Llama-3.2-1B
+tokenizer: meta-llama/Llama-3.2-1B-Instruct
+num-experts: 4
+top-k-experts: 1
+jitter-noise: 0
+use-router: True
+mask-input: True
+max-length: 8192
+
+trainable:
+  - model

models/micro_llama.py

@@ -0,0 +1,588 @@
+from typing import Optional, Tuple, Union, List, Callable
+import logging 
+import yaml
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.distributed as dist
+
+# from transformers.utils import TransformerKwargs
+from transformers import LlamaConfig, AutoConfig, AutoTokenizer, AutoModelForCausalLM
+from transformers.modeling_attn_mask_utils import AttentionMaskConverter
+from transformers.models.llama.modeling_llama import (
+    LlamaRotaryEmbedding, 
+    LlamaRMSNorm, 
+    LlamaMLP,
+    LlamaDecoderLayer,
+    LlamaPreTrainedModel, 
+    GenerationMixin,
+    apply_rotary_pos_emb,
+    eager_attention_forward,
+
+)
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
+from transformers.cache_utils import Cache, StaticCache, DynamicCache
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.processing_utils import Unpack
+from transformers.utils import is_torchdynamo_compiling
+from models.modules import CausalLMOutputWithPast
+from transformers.modeling_layers import GradientCheckpointingLayer
+
+logger = logging.getLogger(__name__)
+
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    min_dtype: float,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """
+    Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+    `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+    Args:
+        attention_mask (`torch.Tensor`):
+            A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+        sequence_length (`int`):
+            The sequence length being processed.
+        target_length (`int`):
+            The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+        dtype (`torch.dtype`):
+            The dtype to use for the 4D attention mask.
+        device (`torch.device`):
+            The device to plcae the 4D attention mask on.
+        min_dtype (`float`):
+            The minimum value representable with the dtype `dtype`.
+        cache_position (`torch.Tensor`):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        batch_size (`torch.Tensor`):
+            Batch size.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask
+    else:
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask
+
+class MiCRoLlamaConfig(LlamaConfig):
+    model_type = "micro_llama"
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.num_experts = kwargs.get("num_experts", 4)
+        self.use_router = kwargs.get("use_router", True)
+        self.num_experts_per_tok = kwargs.get("num_experts_per_tok", 2)
+        self.jitter_noise = kwargs.get("jitter_noise", 0.0)
+        self.loss_method = kwargs.get("loss_method", "all")
+        self.config_path = kwargs.get("config_path", None)
+        
+class MiCRoLlamaDecoderLayer(nn.Module):
+    def __init__(self, config: MiCRoLlamaConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_dim = config.hidden_size
+        self.ffn_dim = config.intermediate_size
+        self.num_experts = config.num_experts
+        self.top_k = config.num_experts_per_tok
+        self.use_router = config.use_router
+        self.ablate = config.ablate
+        self.num_key_value_heads = config.num_key_value_heads
+        self.head_dim = self.hidden_dim // config.num_attention_heads
+        self.gradient_checkpointing = config.gradient_checkpointing
+        if isinstance(self.ablate, str):
+            self.ablate = [self.ablate]
+
+        self.gate = nn.Sequential(
+            nn.Linear(self.hidden_dim, self.hidden_dim, bias=False),
+            nn.Linear(self.hidden_dim, self.num_experts, bias=False)
+        )
+
+        self.num_layers = config.backbone_num_layers
+        self.layer_idx = layer_idx
+
+        self.experts = nn.ModuleList([LlamaDecoderLayer(config, layer_idx * self.num_experts + expert_idx) for expert_idx in range(self.num_experts)])
+
+        self.jitter_noise = config.jitter_noise
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        routing_weights: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        ablate: Optional[List[str]] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+    
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+
+        if ablate is not None:
+            self.ablate = ablate
+
+        if self.training and self.jitter_noise > 0:
+            hidden_states *= torch.empty_like(hidden_states).uniform_(1.0 - self.jitter_noise, 1.0 + self.jitter_noise)
+        
+        if self.use_router:
+            router_logits = self.gate(hidden_states)
+            if "logic" in self.ablate:
+                router_logits[..., 0] = -torch.inf
+            if "social" in self.ablate:
+                router_logits[..., 1] = -torch.inf
+            if "world" in self.ablate:
+                router_logits[..., 2] = -torch.inf
+            if "language" in self.ablate:
+                router_logits[..., 3] = -torch.inf
+            routing_weights = F.softmax(router_logits, dim=-1, dtype=torch.float)
+        else:
+            if len(routing_weights.shape) == 2:
+                routing_weights = routing_weights.unsqueeze(1).tile((1,sequence_length,1)).float()
+            else:
+                routing_weights = routing_weights.float()
+            router_logits = routing_weights
+
+        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+        routing_weights /= (routing_weights.sum(dim=-1, keepdim=True) + 1e-9)
+
+        # we cast back to the input dtype
+        routing_weights = routing_weights.to(hidden_states.dtype)
+
+        # We'll accumulate outputs here
+        final_hidden_states = torch.zeros_like(hidden_states)
+
+        # Flatten final_hidden_states to [batch_size * seq_len, hidden_dim]
+        # so we can do a 2D "index_add_" at the end of each loop.
+        final_hidden_states_2d = final_hidden_states.view(-1, hidden_dim)
+    
+        # One hot encode the selected experts to create an expert mask
+        # this will be used to easily index which expert is going to be sollicitated
+        expert_mask = F.one_hot(selected_experts, num_classes=self.num_experts)
+        #^ [batch_size, seq_len, top_k, num_experts]
+
+        # Loop over all available experts in the model and perform the computation on each expert
+        for expert_idx in range(self.num_experts):
+            expert_layer: LlamaDecoderLayer = self.experts[expert_idx]
+            batch_indices, seq_indices, top_k_indices = torch.where(expert_mask[..., expert_idx])
+
+            if not self.training and sequence_length == 1 and batch_indices.numel() == 0:
+                if past_key_value is not None:
+                    
+                    hidden_state_ln_norm = expert_layer.input_layernorm(hidden_states)
+
+                    input_shape = hidden_state_ln_norm.shape[:-1]
+                    hidden_shape = (*input_shape, -1, self.head_dim)
+
+                    # query_states = expert_layer.self_attn.q_proj(hidden_state_ln_norm).view(hidden_shape).transpose(1, 2)
+                    key_states = expert_layer.self_attn.k_proj(hidden_state_ln_norm).view(hidden_shape).transpose(1, 2)
+                    value_states = expert_layer.self_attn.v_proj(hidden_state_ln_norm).view(hidden_shape).transpose(1, 2)
+
+                    cos, sin = position_embeddings
+                    _, key_states = apply_rotary_pos_emb(key_states, key_states, cos, sin)
+                    # sin and cos are specific to RoPE models; cache_position needed for the static cache
+                    cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+                    past_key_value.update(key_states, value_states, self.layer_idx * self.num_experts + expert_idx, cache_kwargs)
+
+                continue
+            
+            if self.gradient_checkpointing and self.training:
+                current_hidden_states = self._gradient_checkpointing_func(
+                    expert_layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    position_ids,
+                    past_key_value,
+                    output_attentions,
+                    use_cache,
+                    cache_position,
+                    position_embeddings,
+                )[0]
+            else:
+                current_hidden_states = expert_layer(
+                    hidden_states=hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                    position_embeddings=position_embeddings,
+                    **kwargs,
+                )[0]
+            
+            
+            flat_idx = batch_indices * sequence_length + seq_indices
+            expert_weights = routing_weights[batch_indices, seq_indices, top_k_indices].unsqueeze(-1)
+            current_hidden_states = current_hidden_states[batch_indices, seq_indices] * expert_weights
+
+            final_hidden_states_2d.index_add_(0, flat_idx, current_hidden_states.to(hidden_states.dtype))
+       
+        final_hidden_states = final_hidden_states_2d.view(batch_size, sequence_length, hidden_dim)
+        return final_hidden_states, router_logits
+    
+class MiCRoLlama(LlamaPreTrainedModel, GenerationMixin):
+    config_class = MiCRoLlamaConfig
+    def __init__(self, config: MiCRoLlamaConfig):
+        with open(config.config_path, 'r', encoding="utf-8") as file:
+            run_config = yaml.load(file.read(), Loader=yaml.FullLoader)
+
+        self.config: MiCRoLlamaConfig = config
+        self.config.torch_dtype = torch.bfloat16
+        self.config.use_bfloat16 = True
+        self.config._attn_implementation = "flash_attention_2" # {sdpa, flash_attention_2, eager}
+        self.config.backbone_num_layers = self.config.num_hidden_layers
+        self.config.num_hidden_layers = self.config.num_hidden_layers * run_config["num-experts"]
+        self.config.loss_type = "ForCausalLMLoss"
+        
+        super(MiCRoLlama, self).__init__(self.config)
+        self.build_model(run_config)
+
+    def build_model(self, run_config):
+    
+        self.gradient_checkpointing = False
+        self.config.num_experts = run_config["num-experts"]
+        self.config.use_router = run_config["use-router"]
+        self.config.num_experts_per_tok = run_config["top-k-experts"]
+        print(f">> Number of Experts per Token: {self.config.num_experts_per_tok}")
+        self.config.jitter_noise = run_config["jitter-noise"]
+        self.config.loss_method = run_config.get("loss", "all")
+        self.config.gradient_checkpointing = run_config.get("gradient-checkpointing", False)
+        print(f">> Gradient Checkpointing: {self.config.gradient_checkpointing}")
+
+        self.run_config = run_config
+        self.padding_idx = 2 if "smollm2" in run_config["model"] else 128004
+        
+        # MiCRoLlama model
+        self.embed_tokens = nn.Embedding(self.config.vocab_size, self.config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([MiCRoLlamaDecoderLayer(self.config, layer_idx) for layer_idx in range(self.config.backbone_num_layers)])
+        self.lm_head = nn.Linear(self.config.hidden_size, self.config.vocab_size, bias=False)
+        self.rotary_emb = LlamaRotaryEmbedding(config=self.config)
+        self.final_norm = LlamaRMSNorm(self.config.hidden_size, eps=self.config.rms_norm_eps)
+
+        if "model" not in run_config["trainable"]:
+            print(">> Freezing Model Except Routing Gates")
+            for param in self.parameters():
+                param.requires_grad = False
+
+            for layer in self.layers:
+                layer: MiCRoLlamaDecoderLayer
+                for param in layer.gate.parameters():
+                    param.requires_grad = True
+
+        if "experts-router" not in run_config["trainable"]:
+            print(">> Freezing Routing Gates")
+            for layer in self.layers:
+                layer: MiCRoLlamaDecoderLayer
+                for param in layer.gate.parameters():
+                    param.requires_grad = False
+
+
+
+    def forward(self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        experts_ablate: Optional[List[str]] = None,
+        routing_weights: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ):
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+        )
+
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+
+        all_routing_weights = ()
+ 
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training and False:
+                layer_outputs, router_logits = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    routing_weights,
+                    causal_mask,
+                    position_ids,
+                    experts_ablate,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                    cache_position,
+                    position_embeddings,
+                )
+            else:
+                layer_outputs, router_logits = decoder_layer(
+                    hidden_states,
+                    routing_weights=routing_weights,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    ablate=experts_ablate,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                    position_embeddings=position_embeddings,
+                    **kwargs,
+                )
+
+            hidden_states = layer_outputs
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+            all_routing_weights += (router_logits,)
+
+        hidden_states = self.final_norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+    
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+
+        if not return_dict:
+            output = (logits,) + (past_key_values, all_hidden_states, all_self_attns, all_routing_weights) if use_cache else (logits, all_hidden_states, all_self_attns, all_routing_weights) 
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=past_key_values if use_cache else None,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            routing_weights=all_routing_weights,
+        )
+
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool,
+    ):
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and 0.0 in attention_mask:
+                return attention_mask
+            return None
+
+        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+        # to infer the attention mask.
+        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+        using_static_cache = isinstance(past_key_values, StaticCache)
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                is_training=self.training,
+            ):
+                return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        min_dtype = torch.finfo(dtype).min
+        sequence_length = input_tensor.shape[1]
+        if using_static_cache:
+            target_length = past_key_values.get_max_length()
+        else:
+            target_length = (
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
+            )
+
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            min_dtype=min_dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type == "cuda"
+            and not output_attentions
+        ):
+            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+            # Details: https://github.com/pytorch/pytorch/issues/110213
+            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+        return causal_mask
+
+    def load_pretrained(self, model_name):
+        base_model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16)
+        self.lm_head.load_state_dict(base_model.lm_head.state_dict())
+        self.embed_tokens.load_state_dict(base_model.get_input_embeddings().state_dict())
+        self.rotary_emb.load_state_dict(base_model.model.rotary_emb.state_dict())
+        self.final_norm.load_state_dict(base_model.model.norm.state_dict())
+        for layer_idx, layer in enumerate(self.layers):
+            base_model_layer = base_model.model.layers[layer_idx].state_dict()
+            for expert in layer.experts:
+                expert.load_state_dict(base_model_layer)
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        cache_position=None,
+        position_ids=None,
+        experts_ablate=None,
+        use_cache=True,
+        num_logits_to_keep=None,
+        **kwargs,
+    ):
+        
+        # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
+        # Exception 1: when passing input_embeds, input_ids may be missing entries
+        # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
+        if past_key_values is not None:
+            if inputs_embeds is not None:  # Exception 1
+                input_ids = input_ids[:, -cache_position.shape[0] :]
+            elif input_ids.shape[1] != cache_position.shape[0]:  # Default case (the "else", a no op, is Exception 2)
+                input_ids = input_ids[:, cache_position]
+
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -input_ids.shape[1] :]
+
+                # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s  `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture.
+                position_ids = position_ids.clone(memory_format=torch.contiguous_format)
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and cache_position[0] == 0:
+            model_inputs = {"inputs_embeds": inputs_embeds, "input_ids": None}
+        else:
+            # The clone here is for the same reason as for `position_ids`.
+            model_inputs = {"input_ids": input_ids.clone(memory_format=torch.contiguous_format), "inputs_embeds": None}
+
+        if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2:
+            if model_inputs["inputs_embeds"] is not None:
+                batch_size, sequence_length, _ = model_inputs["inputs_embeds"].shape
+                device = model_inputs["inputs_embeds"].device
+            else:
+                batch_size, sequence_length = model_inputs["input_ids"].shape
+                device = model_inputs["input_ids"].device
+
+            dtype = self.lm_head.weight.dtype
+            min_dtype = torch.finfo(dtype).min
+
+            attention_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+                attention_mask,
+                sequence_length=sequence_length,
+                target_length=past_key_values.get_max_length(),
+                dtype=dtype,
+                device=device,
+                min_dtype=min_dtype,
+                cache_position=cache_position,
+                batch_size=batch_size,
+            )
+
+        if num_logits_to_keep is not None:
+            model_inputs["num_logits_to_keep"] = num_logits_to_keep
+
+        model_inputs.update(
+            {
+                "experts_ablate": experts_ablate,
+                "position_ids": position_ids,
+                "cache_position": cache_position,
+                "past_key_values": past_key_values,
+                "use_cache": use_cache,
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+
+AutoConfig.register("micro_llama", MiCRoLlamaConfig)
+AutoModelForCausalLM.register(MiCRoLlamaConfig, MiCRoLlama)

models/micro_moe_llama.py

@@ -0,0 +1,725 @@
+from typing import Optional, Tuple, Union, List, Callable
+import logging 
+import yaml
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.distributed as dist
+
+from transformers import LlamaConfig, AutoModelForCausalLM, AutoConfig
+from transformers.modeling_attn_mask_utils import AttentionMaskConverter
+from transformers.models.llama.modeling_llama import (
+    LlamaRotaryEmbedding, 
+    LlamaRMSNorm, 
+    LlamaMLP,
+    LlamaAttention,
+    LlamaForCausalLM,
+    LlamaPreTrainedModel, 
+    GenerationMixin,
+    apply_rotary_pos_emb,
+    eager_attention_forward,
+
+)
+from transformers.modeling_layers import GradientCheckpointingLayer
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
+from transformers.cache_utils import Cache, StaticCache, DynamicCache
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.processing_utils import Unpack
+from transformers.utils import is_torchdynamo_compiling
+from transformers.activations import ACT2FN
+from models.modules import CausalLMOutputWithPast
+
+logger = logging.getLogger(__name__)
+
+def keep_alive_zero(model):
+    z = 0.0
+    for p in model.parameters():
+        if p.requires_grad:
+            # one scalar per param to avoid heavy sums
+            z = z + (p.view(-1)[0] * 0.0)
+    return z
+
+class MiCRoLlamaMoEConfig(LlamaConfig):
+    model_type = "micro_llama_moe"
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.num_experts = kwargs.get("num_experts", 4)
+        self.use_router = kwargs.get("use_router", True)
+        self.num_experts_per_tok = kwargs.get("num_experts_per_tok", 2)
+        self.jitter_noise = kwargs.get("jitter_noise", 0.0)
+        self.loss_method = kwargs.get("loss_method", "all")
+        self.config_path = kwargs.get("config_path", None)
+
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    min_dtype: float,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """
+    Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+    `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+    Args:
+        attention_mask (`torch.Tensor`):
+            A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+        sequence_length (`int`):
+            The sequence length being processed.
+        target_length (`int`):
+            The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+        dtype (`torch.dtype`):
+            The dtype to use for the 4D attention mask.
+        device (`torch.device`):
+            The device to plcae the 4D attention mask on.
+        min_dtype (`float`):
+            The minimum value representable with the dtype `dtype`.
+        cache_position (`torch.Tensor`):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        batch_size (`torch.Tensor`):
+            Batch size.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask
+    else:
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask
+
+class DummyModule(nn.Module):
+    def __init__(self):
+        super().__init__()
+    def forward(self, x):
+        return x
+    
+class LlamaSparseMiCRoMoEBlock(nn.Module):
+    """
+    This implementation is
+    strictly equivalent to standard MoE with full capacity (no
+    dropped tokens). It's faster since it formulates MoE operations
+    in terms of block-sparse operations to accommodate imbalanced
+    assignments of tokens to experts, whereas standard MoE either
+    (1) drop tokens at the cost of reduced performance or (2) set
+    capacity factor to number of experts and thus waste computation
+    and memory on padding.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.hidden_dim = config.hidden_size
+        self.ffn_dim = config.intermediate_size
+        self.num_experts = config.num_experts
+        self.top_k = config.num_experts_per_tok
+        self.use_router = config.use_router
+        self.ablate = config.ablate
+
+        # gating
+        self.gate = nn.Sequential(
+            nn.Linear(self.hidden_dim, self.hidden_dim, bias=False),
+            nn.Linear(self.hidden_dim, self.num_experts, bias=False)
+        )
+
+        self.experts = nn.ModuleList([LlamaMLP(config) for _ in range(self.num_experts)])
+
+        self.dummy = DummyModule()
+
+        # Jitter parameters
+        self.jitter_noise = config.jitter_noise
+
+    def forward(self, hidden_states: torch.Tensor, routing_weights: Optional[torch.Tensor] = None) -> torch.Tensor:
+        """ """
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+        if self.training and self.jitter_noise > 0:
+            hidden_states *= torch.empty_like(hidden_states).uniform_(1.0 - self.jitter_noise, 1.0 + self.jitter_noise)
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        
+        if self.use_router:
+            router_logits = self.gate(hidden_states)
+            if "logic" in self.ablate:
+                router_logits[..., 0] = -torch.inf
+            if "social" in self.ablate:
+                router_logits[..., 1] = -torch.inf
+            if "world" in self.ablate:
+                router_logits[..., 2] = -torch.inf
+            if "language" in self.ablate:
+                router_logits[..., 3] = -torch.inf
+            routing_weights = F.softmax(router_logits, dim=-1, dtype=torch.float)
+        else:
+            routing_weights = routing_weights.reshape(-1, 4).float()
+            router_logits = routing_weights
+        # router_logits: (batch * sequence_length, n_experts)
+        
+        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+        routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
+        # we cast back to the input dtype
+        routing_weights = routing_weights.to(hidden_states.dtype)
+
+        final_hidden_states = torch.zeros(
+            (batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device
+        )
+
+        H_up = self.experts[0].up_proj.out_features
+        Y_up = hidden_states.new_zeros((batch_size, sequence_length, self.num_experts, H_up))
+
+
+        # One hot encode the selected experts to create an expert mask
+        # this will be used to easily index which expert is going to be sollicitated
+        expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
+
+        expert_hitted = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero()
+        for expert_idx in expert_hitted:
+            expert_layer = self.experts[expert_idx]
+            idx, top_x = torch.where(expert_mask[expert_idx].squeeze(0))
+            # Index the correct hidden states and compute the expert hidden state for
+            # the current expert. We need to make sure to multiply the output hidden
+            # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
+            current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
+
+            # --- Hook to capture up-proj output BEFORE nonlinearity ---
+            captured_up = []
+            def _up_hook(m, inp, out):
+                # out shape: [N_e, H_up]
+                captured_up.append(out.detach())
+
+            h = expert_layer.up_proj.register_forward_hook(_up_hook)
+
+            current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None]
+            h.remove()
+
+            # Scatter captured up-proj per-token into Y_up[b, t, expert, :]
+            if captured_up:
+                up = captured_up[-1]  # [N_e, H_up]
+                b_idx = top_x // sequence_length
+                t_idx = top_x %  sequence_length
+                # Y_up[b,t,e,:] = up[n,:]
+                Y_up[b_idx, t_idx, expert_idx, :] = up
+            
+            # However `index_add_` only support torch tensors for indexing so we'll use
+            # the `top_x` tensor here.
+            final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
+        final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
+
+        self.dummy(Y_up)
+        return final_hidden_states, router_logits
+
+class LlamaMiCRoMoEDecoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: MiCRoLlamaMoEConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.self_attn = LlamaAttention(config=config, layer_idx=layer_idx)
+
+        self.block_sparse_moe = LlamaSparseMiCRoMoEBlock(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)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        routing_weights: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[tuple[torch.Tensor]] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> torch.FloatTensor:
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, _ = self.self_attn(
+            hidden_states=hidden_states,
+            position_embeddings=position_embeddings,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            cache_position=cache_position,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states, router_logits = self.block_sparse_moe(hidden_states, routing_weights)
+        hidden_states = residual + hidden_states
+
+        return hidden_states, router_logits
+
+    
+class MiCRoLlamaMoE(LlamaPreTrainedModel, GenerationMixin):
+    config_class = MiCRoLlamaMoEConfig
+    def __init__(self, config):
+        with open(config.config_path, 'r', encoding="utf-8") as file:
+            run_config = yaml.load(file.read(), Loader=yaml.FullLoader)
+
+        self.config: MiCRoLlamaMoEConfig = config
+        self.config.torch_dtype = torch.bfloat16
+        self.config.use_bfloat16 = True
+        self.config._attn_implementation = "flash_attention_2" # {sdpa, flash_attention_2, eager}
+        self.config.use_cache = True
+        self.config.backbone_num_layers = self.config.num_hidden_layers
+        self.config.num_hidden_layers = self.config.num_hidden_layers
+        self.config.loss_type = "ForCausalLMLoss"
+
+        super(MiCRoLlamaMoE, self).__init__(self.config)
+        self.build_model(run_config)
+
+    def build_model(self, run_config):
+    
+        self.config.num_experts = run_config["num-experts"]
+        self.config.use_router = run_config["use-router"]
+        self.config.num_experts_per_tok = run_config["top-k-experts"]
+        print(f">> Top-K Experts Per Token: {self.config.num_experts_per_tok}")
+        self.config.jitter_noise = run_config["jitter-noise"]
+        self.config.loss_method = run_config.get("loss", "all")
+        self.router_aux_loss_coef = run_config["router-aux-loss-coef"]
+        self.use_load_balancing = run_config.get("use-load-balancing", False)
+
+        self.config.gradient_checkpointing = run_config.get("gradient-checkpointing", False)
+        self.gradient_checkpointing = self.config.gradient_checkpointing
+
+        print(f">> Gradient Checkpointing: {self.config.gradient_checkpointing}")
+
+        self.run_config = run_config
+        self.padding_idx = 2 if "smollm2" in run_config["model"] else 128004
+        
+        # LlamaMoE model
+        self.embed_tokens = nn.Embedding(self.config.vocab_size, self.config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([LlamaMiCRoMoEDecoderLayer(self.config, layer_idx) for layer_idx in range(self.config.backbone_num_layers)])
+        self.lm_head = nn.Linear(self.config.hidden_size, self.config.vocab_size, bias=False)
+        self.rotary_emb = LlamaRotaryEmbedding(config=self.config)
+        self.final_norm = LlamaRMSNorm(self.config.hidden_size, eps=self.config.rms_norm_eps)
+        
+        if "model" not in run_config["trainable"]:
+            print(">> Freezing Model Except Experts + Routing Gates")
+            for param in self.parameters():
+                param.requires_grad = False
+
+            for layer in self.layers:
+                layer: LlamaMiCRoMoEDecoderLayer
+                for param in layer.block_sparse_moe.parameters():
+                    param.requires_grad = True
+
+        if "experts" not in run_config["trainable"]:
+            print(">> Freezing Experts")
+            for layer in self.layers:
+                layer: LlamaMiCRoMoEDecoderLayer
+                for param in layer.block_sparse_moe.experts.parameters():
+                    param.requires_grad = False
+
+        if "experts-router" not in run_config["trainable"]:
+            print(">> Freezing Routing Gates")
+            for layer in self.layers:
+                layer: LlamaMiCRoMoEDecoderLayer
+                for param in layer.block_sparse_moe.gate.parameters():
+                    param.requires_grad = False
+
+
+    def forward(self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        experts_ablate: Optional[List[str]] = None,
+        routing_weights: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ):
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+        )
+
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+
+        all_routing_weights = ()
+
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs, router_logits = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    position_embeddings,
+                    routing_weights,
+                    causal_mask,
+                    position_ids,
+                    past_key_values,
+                    cache_position,
+                )
+            else:
+                layer_outputs, router_logits = decoder_layer(
+                    hidden_states,
+                    position_embeddings=position_embeddings,
+                    routing_weights=routing_weights,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                    **kwargs,
+                )
+
+            hidden_states = layer_outputs
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+            all_routing_weights += (router_logits,)
+
+        hidden_states = self.final_norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+    
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+
+            loss += keep_alive_zero(self)
+            
+            aux_loss = None
+            if self.use_load_balancing:
+                aux_loss = load_balancing_loss_func(
+                    all_routing_weights,
+                    self.config.num_experts,
+                    self.config.num_experts_per_tok,
+                    attention_mask,
+                )
+                loss += self.router_aux_loss_coef * aux_loss.to(loss.device)  # make sure to reside in the same device
+
+        if not return_dict:
+            output = (logits,) + (past_key_values, all_hidden_states, all_self_attns, all_routing_weights) if use_cache else (logits, all_hidden_states, all_self_attns, all_routing_weights) 
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=past_key_values if use_cache else None,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            routing_weights=all_routing_weights,
+        )
+
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool,
+    ):
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and 0.0 in attention_mask:
+                return attention_mask
+            return None
+
+        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+        # to infer the attention mask.
+        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+        using_static_cache = isinstance(past_key_values, StaticCache)
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                is_training=self.training,
+            ):
+                return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        min_dtype = torch.finfo(dtype).min
+        sequence_length = input_tensor.shape[1]
+        if using_static_cache:
+            target_length = past_key_values.get_max_length()
+        else:
+            target_length = (
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
+            )
+
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            min_dtype=min_dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type == "cuda"
+            and not output_attentions
+        ):
+            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+            # Details: https://github.com/pytorch/pytorch/issues/110213
+            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+        return causal_mask
+
+    def load_pretrained(self, model_name):
+        base_model: LlamaForCausalLM = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16)
+        self.lm_head.load_state_dict(base_model.lm_head.state_dict())
+        self.embed_tokens.load_state_dict(base_model.get_input_embeddings().state_dict())
+        self.rotary_emb.load_state_dict(base_model.model.rotary_emb.state_dict())
+        self.final_norm.load_state_dict(base_model.model.norm.state_dict())
+
+        for layer_idx, layer in enumerate(self.layers):
+            
+            attn_layer = base_model.model.layers[layer_idx].self_attn.state_dict()
+            layer.self_attn.load_state_dict(attn_layer)
+            
+            input_layernorm = base_model.model.layers[layer_idx].input_layernorm.state_dict()
+            layer.input_layernorm.load_state_dict(input_layernorm)
+
+            post_attention_layernorm = base_model.model.layers[layer_idx].post_attention_layernorm.state_dict()
+            layer.post_attention_layernorm.load_state_dict(post_attention_layernorm)
+            
+            mlp_model_layer = base_model.model.layers[layer_idx].mlp.state_dict()
+            for expert in layer.block_sparse_moe.experts:
+                expert.load_state_dict(mlp_model_layer)
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        cache_position=None,
+        position_ids=None,
+        experts_ablate=None,
+        use_cache=True,
+        num_logits_to_keep=None,
+        **kwargs,
+    ):
+        
+        # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
+        # Exception 1: when passing input_embeds, input_ids may be missing entries
+        # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
+        if past_key_values is not None:
+            if inputs_embeds is not None:  # Exception 1
+                input_ids = input_ids[:, -cache_position.shape[0] :]
+            elif input_ids.shape[1] != cache_position.shape[0]:  # Default case (the "else", a no op, is Exception 2)
+                input_ids = input_ids[:, cache_position]
+
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -input_ids.shape[1] :]
+
+                # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s  `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture.
+                position_ids = position_ids.clone(memory_format=torch.contiguous_format)
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and cache_position[0] == 0:
+            model_inputs = {"inputs_embeds": inputs_embeds, "input_ids": None}
+        else:
+            # The clone here is for the same reason as for `position_ids`.
+            model_inputs = {"input_ids": input_ids.clone(memory_format=torch.contiguous_format), "inputs_embeds": None}
+
+        if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2:
+            if model_inputs["inputs_embeds"] is not None:
+                batch_size, sequence_length, _ = model_inputs["inputs_embeds"].shape
+                device = model_inputs["inputs_embeds"].device
+            else:
+                batch_size, sequence_length = model_inputs["input_ids"].shape
+                device = model_inputs["input_ids"].device
+
+            dtype = self.lm_head.weight.dtype
+            min_dtype = torch.finfo(dtype).min
+
+            attention_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+                attention_mask,
+                sequence_length=sequence_length,
+                target_length=past_key_values.get_max_length(),
+                dtype=dtype,
+                device=device,
+                min_dtype=min_dtype,
+                cache_position=cache_position,
+                batch_size=batch_size,
+            )
+
+        if num_logits_to_keep is not None:
+            model_inputs["num_logits_to_keep"] = num_logits_to_keep
+
+        model_inputs.update(
+            {
+                "experts_ablate": experts_ablate,
+                "position_ids": position_ids,
+                "cache_position": cache_position,
+                "past_key_values": past_key_values,
+                "use_cache": use_cache,
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+    
+def load_balancing_loss_func(
+    gate_logits: Union[torch.Tensor, tuple[torch.Tensor], None],
+    num_experts: Optional[int] = None,
+    top_k=2,
+    attention_mask: Optional[torch.Tensor] = None,
+) -> Union[torch.Tensor, int]:
+    r"""
+    Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.
+
+    See Switch Transformer (https://huggingface.co/papers/2101.03961) for more details. This function implements the loss
+    function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between
+    experts is too unbalanced.
+
+    Args:
+        gate_logits:
+            Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of
+            shape [batch_size X sequence_length, num_experts].
+        num_experts:
+            Number of experts
+        top_k:
+            The number of experts to route per-token, can be also interpreted as the `top-k` routing
+            parameter.
+        attention_mask (`torch.Tensor`, *optional*):
+            The attention_mask used in forward function
+            shape [batch_size X sequence_length] if not None.
+
+    Returns:
+        The auxiliary loss.
+    """
+    if gate_logits is None or not isinstance(gate_logits, tuple):
+        return 0
+
+    if isinstance(gate_logits, tuple):
+        compute_device = gate_logits[0].device
+        concatenated_gate_logits = torch.cat([layer_gate.to(compute_device) for layer_gate in gate_logits], dim=0)
+
+    routing_weights = torch.nn.functional.softmax(concatenated_gate_logits, dim=-1)
+
+    _, selected_experts = torch.topk(routing_weights, top_k, dim=-1)
+
+    expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts)
+
+    if attention_mask is None:
+        # Compute the percentage of tokens routed to each experts
+        tokens_per_expert = torch.mean(expert_mask.float(), dim=0)
+
+        # Compute the average probability of routing to these experts
+        router_prob_per_expert = torch.mean(routing_weights, dim=0)
+    else:
+        batch_size, sequence_length = attention_mask.shape
+        num_hidden_layers = concatenated_gate_logits.shape[0] // (batch_size * sequence_length)
+
+        # Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask
+        expert_attention_mask = (
+            attention_mask[None, :, :, None, None]
+            .expand((num_hidden_layers, batch_size, sequence_length, top_k, num_experts))
+            .reshape(-1, top_k, num_experts)
+            .to(compute_device)
+        )
+
+        # Compute the percentage of tokens routed to each experts
+        tokens_per_expert = torch.sum(expert_mask.float() * expert_attention_mask, dim=0) / torch.sum(
+            expert_attention_mask, dim=0
+        )
+
+        # Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert
+        router_per_expert_attention_mask = (
+            attention_mask[None, :, :, None]
+            .expand((num_hidden_layers, batch_size, sequence_length, num_experts))
+            .reshape(-1, num_experts)
+            .to(compute_device)
+        )
+
+        # Compute the average probability of routing to these experts
+        router_prob_per_expert = torch.sum(routing_weights * router_per_expert_attention_mask, dim=0) / torch.sum(
+            router_per_expert_attention_mask, dim=0
+        )
+
+    overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0))
+    return overall_loss * num_experts
+
+
+AutoConfig.register("micro_llama_moe", MiCRoLlamaMoEConfig)
+AutoModelForCausalLM.register(MiCRoLlamaMoEConfig, MiCRoLlamaMoE)

models/micro_olmo.py

@@ -0,0 +1,528 @@
+from typing import Callable, Optional, Tuple, Union
+
+import yaml
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+
+from transformers import AutoModelForCausalLM
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.generation import GenerationMixin
+from transformers.modeling_attn_mask_utils import AttentionMaskConverter
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+# from transformers.modeling_layers import GradientCheckpointingLayer
+from transformers.modeling_outputs import BaseModelOutputWithPast
+from transformers.processing_utils import Unpack
+from transformers.utils import (
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_torch_flex_attn_available,
+    logging,
+    replace_return_docstrings,
+)
+from transformers.models.olmo2.configuration_olmo2 import Olmo2Config
+from transformers.models.olmo2.modeling_olmo2 import (
+    Olmo2RMSNorm,
+    Olmo2Attention,
+    Olmo2MLP,
+    Olmo2DecoderLayer,
+    Olmo2RotaryEmbedding,
+    Olmo2PreTrainedModel,
+    rotate_half,
+    apply_rotary_pos_emb,
+    repeat_kv,
+    eager_attention_forward,
+)
+
+
+if is_torch_flex_attn_available():
+    from torch.nn.attention.flex_attention import BlockMask
+
+from models.modules import CausalLMOutputWithPast
+
+logger = logging.get_logger(__name__)
+
+class MiCRoOLMo2DecoderLayer(nn.Module):
+    def __init__(self, config: Olmo2Config, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.num_experts  = config.num_experts
+        self.top_k        = config.num_experts_per_tok
+        self.use_router   = config.use_router
+        self.ablate       = config.ablate or []
+        self.num_layers   = config.backbone_num_layers
+        self.layer_idx    = layer_idx
+        self.jitter_noise = config.jitter_noise
+        self.config = config
+        self.head_dim = config.hidden_size // config.num_attention_heads
+
+        if isinstance(self.ablate, str):
+            self.ablate = [self.ablate]
+
+        # gating head
+        self.gate = nn.Sequential(
+            nn.Linear(self.hidden_size, self.hidden_size, bias=False),
+            nn.Linear(self.hidden_size, self.num_experts, bias=False),
+        )
+
+        self.experts = nn.ModuleList([
+            Olmo2DecoderLayer(config, layer_idx * self.num_experts + expert_idx)
+            for expert_idx in range(self.num_experts)
+        ])
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        routing_weights: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+    
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+
+        if self.training and self.jitter_noise > 0:
+            hidden_states *= torch.empty_like(hidden_states).uniform_(1.0 - self.jitter_noise, 1.0 + self.jitter_noise)
+        
+        if self.use_router:
+            router_logits = self.gate(hidden_states)
+            if "logic" in self.ablate:
+                router_logits[..., 0] = -torch.inf
+            if "social" in self.ablate:
+                router_logits[..., 1] = -torch.inf
+            if "world" in self.ablate:
+                router_logits[..., 2] = -torch.inf
+            if "language" in self.ablate:
+                router_logits[..., 3] = -torch.inf
+            routing_weights = F.softmax(router_logits, dim=-1, dtype=torch.float)
+        else:
+            if len(routing_weights.shape) == 2:
+                routing_weights = routing_weights.unsqueeze(1).tile((1,sequence_length,1)).float()
+            else:
+                routing_weights = routing_weights.float()
+            router_logits = routing_weights
+
+        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+        routing_weights /= (routing_weights.sum(dim=-1, keepdim=True) + 1e-9)
+
+        # we cast back to the input dtype
+        routing_weights = routing_weights.to(hidden_states.dtype)
+
+        # We'll accumulate outputs here
+        final_hidden_states = torch.zeros_like(hidden_states)
+
+        # Flatten final_hidden_states to [batch_size * seq_len, hidden_dim]
+        # so we can do a 2D "index_add_" at the end of each loop.
+        final_hidden_states_2d = final_hidden_states.view(-1, hidden_dim)
+    
+        # One hot encode the selected experts to create an expert mask
+        # this will be used to easily index which expert is going to be sollicitated
+        expert_mask = F.one_hot(selected_experts, num_classes=self.num_experts)
+        #^ [batch_size, seq_len, top_k, num_experts]
+
+        # Loop over all available experts in the model and perform the computation on each expert
+        for expert_idx in range(self.num_experts):
+            expert_layer: Olmo2DecoderLayer = self.experts[expert_idx]
+            batch_indices, seq_indices, top_k_indices = torch.where(expert_mask[..., expert_idx])
+        
+            if not self.training and sequence_length == 1 and batch_indices.numel() == 0:
+                if past_key_value is not None:
+                    
+                    input_shape = hidden_states.shape[:-1]
+                    hidden_shape = (*input_shape, -1, self.head_dim)
+
+                    key_states = expert_layer.self_attn.k_proj(hidden_states)
+                    key_states = expert_layer.self_attn.k_norm(key_states).view(hidden_shape).transpose(1, 2)
+                    value_states = expert_layer.self_attn.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+
+                    cos, sin = position_embeddings
+                    _, key_states = apply_rotary_pos_emb(key_states, key_states, cos, sin)
+                    # sin and cos are specific to RoPE models; cache_position needed for the static cache
+                    cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+                    past_key_value.update(key_states, value_states, self.layer_idx * self.num_experts + expert_idx, cache_kwargs)
+
+                continue
+        
+            current_hidden_states = expert_layer(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
+                **kwargs,
+            )[0]
+            
+            flat_idx = batch_indices * sequence_length + seq_indices
+            expert_weights = routing_weights[batch_indices, seq_indices, top_k_indices].unsqueeze(-1)
+            current_hidden_states = current_hidden_states[batch_indices, seq_indices] * expert_weights
+
+            final_hidden_states_2d.index_add_(0, flat_idx, current_hidden_states.to(hidden_states.dtype))
+
+        final_hidden_states = final_hidden_states_2d.view(batch_size, sequence_length, hidden_dim)
+        return final_hidden_states, router_logits
+    
+class MiCRoOLMo(Olmo2PreTrainedModel, GenerationMixin):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Olmo2DecoderLayer`]
+
+    Args:
+        config: Olmo2Config
+    """
+
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config: Olmo2Config):
+        with open(config.config_path, 'r', encoding="utf-8") as file:
+            run_config = yaml.load(file.read(), Loader=yaml.FullLoader)
+
+        self.config: Olmo2Config = config
+        self.config.torch_dtype = torch.bfloat16
+        self.config.use_bfloat16 = True
+        self.config._attn_implementation = "flash_attention_2" # {sdpa, flash_attention_2, eager}
+        self.config.use_cache = True
+        self.config.backbone_num_layers = self.config.num_hidden_layers
+        self.config.num_hidden_layers = self.config.num_hidden_layers * run_config["num-experts"]
+        self.config.loss_type = "ForCausalLMLoss"
+
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.gradient_checkpointing = False
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.build_model(run_config)
+    
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, value):
+        self.lm_head = value
+
+    def build_model(self, run_config):
+        self.gradient_checkpointing = False
+        self.config.num_experts = run_config["num-experts"]
+        self.config.use_router = run_config["use-router"]
+        self.config.num_experts_per_tok = run_config["top-k-experts"]
+        self.config.jitter_noise = run_config["jitter-noise"]
+        self.config.loss_method = run_config.get("loss", "all")
+
+        self.run_config = run_config        
+        # Qwen2 model
+        self.embed_tokens = nn.Embedding(self.config.vocab_size, self.config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([MiCRoOLMo2DecoderLayer(self.config, layer_idx) for layer_idx in range(self.config.backbone_num_layers)])
+        self.lm_head = nn.Linear(self.config.hidden_size, self.config.vocab_size, bias=False)
+        self.rotary_emb = Olmo2RotaryEmbedding(config=self.config)
+        self.norm = Olmo2RMSNorm(self.config.hidden_size, eps=self.config.rms_norm_eps)
+
+        # Freeze Model
+        for param in self.parameters():
+            param.requires_grad = False
+
+        # Unfreeze Modules
+        if "reasoners" in run_config["trainable"]:
+            print(">> Unfreezing Reasoning Modules")
+            for layer in self.layers:
+                layer: MiCRoOLMo2DecoderLayer
+                for param in layer.experts.parameters():
+                    param.requires_grad = True
+
+        if "model" in run_config["trainable"]:
+            print(">> Unfreezing Model")
+            for param in self.layers.parameters():
+                param.requires_grad = True
+
+            for param in self.lm_head.parameters():
+                param.requires_grad = True
+
+            for param in self.rotary_emb.parameters():
+                param.requires_grad = True
+
+            for param in self.norm.parameters():
+                param.requires_grad = True
+
+            for param in self.embed_tokens.parameters():
+                param.requires_grad = True
+
+            for layer in self.layers:
+                for param in layer.gate.parameters():
+                    param.requires_grad = False
+
+
+        if "experts-router" in run_config["trainable"]:
+            print(">> Unfreezing Experts Router")
+            for layer in self.layers:
+                for param in layer.gate.parameters():
+                    param.requires_grad = True
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        routing_weights: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> BaseModelOutputWithPast:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
+
+        # TODO (joao): remove this exception in v4.56 -- it exists for users that try to pass a legacy cache
+        if not isinstance(past_key_values, (type(None), Cache)):
+            raise ValueError("The `past_key_values` should be either a `Cache` object or `None`.")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+        )
+
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_routing_weights = ()
+
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            layer_outputs, router_logits = decoder_layer(
+                hidden_states,
+                routing_weights=routing_weights,
+                attention_mask=causal_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_values,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
+                **kwargs,
+                # **flash_attn_kwargs,
+            )
+
+            hidden_states = layer_outputs
+
+            # if output_attentions:
+            #     all_self_attns += (layer_outputs[1],)
+                
+            all_routing_weights += (router_logits,)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+    
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=past_key_values if use_cache else None,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            routing_weights=all_routing_weights,
+        )
+
+    def load_pretrained(self, model_name):
+        base_model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16)
+        self.lm_head.load_state_dict(base_model.lm_head.state_dict())
+        self.embed_tokens.load_state_dict(base_model.get_input_embeddings().state_dict())
+        self.rotary_emb.load_state_dict(base_model.model.rotary_emb.state_dict())
+        self.norm.load_state_dict(base_model.model.norm.state_dict())
+        for layer_idx, layer in enumerate(self.layers):
+            base_model_layer = base_model.model.layers[layer_idx].state_dict()
+            for expert in layer.experts:
+                expert.load_state_dict(base_model_layer)
+
+    def _update_causal_mask(
+        self,
+        attention_mask: Union[torch.Tensor, "BlockMask"],
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool = False,
+    ):
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and (attention_mask == 0.0).any():
+                return attention_mask
+            return None
+        if self.config._attn_implementation == "flex_attention":
+            if isinstance(attention_mask, torch.Tensor):
+                attention_mask = make_flex_block_causal_mask(attention_mask)
+            return attention_mask
+
+        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+        # to infer the attention mask.
+        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+        using_compilable_cache = past_key_values.is_compileable if past_key_values is not None else False
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if self.config._attn_implementation == "sdpa" and not using_compilable_cache and not output_attentions:
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                is_training=self.training,
+            ):
+                return None
+
+        dtype = input_tensor.dtype
+        sequence_length = input_tensor.shape[1]
+        if using_compilable_cache:
+            target_length = past_key_values.get_max_cache_shape()
+        else:
+            target_length = (
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
+            )
+
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type in ["cuda", "xpu", "npu"]
+            and not output_attentions
+        ):
+            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+            # Details: https://github.com/pytorch/pytorch/issues/110213
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+        return causal_mask
+
+    @staticmethod
+    def _prepare_4d_causal_attention_mask_with_cache_position(
+        attention_mask: torch.Tensor,
+        sequence_length: int,
+        target_length: int,
+        dtype: torch.dtype,
+        cache_position: torch.Tensor,
+        batch_size: int,
+        **kwargs,
+    ):
+        """
+        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+        Args:
+            attention_mask (`torch.Tensor`):
+                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
+                `(batch_size, 1, query_length, key_value_length)`.
+            sequence_length (`int`):
+                The sequence length being processed.
+            target_length (`int`):
+                The target length: when generating with static cache, the mask should be as long as the static cache,
+                to account for the 0 padding, the part of the cache that is not filled yet.
+            dtype (`torch.dtype`):
+                The dtype to use for the 4D attention mask.
+            cache_position (`torch.Tensor`):
+                Indices depicting the position of the input sequence tokens in the sequence.
+            batch_size (`torch.Tensor`):
+                Batch size.
+        """
+        if attention_mask is not None and attention_mask.dim() == 4:
+            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+            causal_mask = attention_mask
+        else:
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = torch.full(
+                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=cache_position.device
+            )
+            if sequence_length != 1:
+                causal_mask = torch.triu(causal_mask, diagonal=1)
+            causal_mask *= torch.arange(target_length, device=cache_position.device) > cache_position.reshape(-1, 1)
+            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+            if attention_mask is not None:
+                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
+                    causal_mask.device
+                )
+                padding_mask = padding_mask == 0
+                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                    padding_mask, min_dtype
+                )
+
+        return causal_mask
+
+
+__all__ = ["MiCRoOLMo"]

models/modules.py

@@ -0,0 +1,42 @@
+from dataclasses import dataclass
+from typing import Optional, Tuple, List, Union
+
+import torch
+from transformers.modeling_outputs import ModelOutput
+
+
+@dataclass
+class CausalLMOutputWithPast(ModelOutput):
+    """
+    Base class for causal language model (or autoregressive) outputs.
+
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+            Language modeling loss (for next-token prediction).
+        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`)
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
+            `past_key_values` input) to speed up sequential decoding.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
+    attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
+    routing_weights: Optional[Tuple[torch.FloatTensor, ...]] = None

requirements.txt

@@ -0,0 +1,3 @@
+gradio>=4.44.0
+plotly>=5.22.0
+pandas>=2.2.0

router_backend.py

@@ -0,0 +1,223 @@
+# router_backend.py
+"""
+Plug your real model routing function here.
+
+Implement the function:
+    get_expert_routing(model_id: str, prompt: str) -> list[float] | dict[str, float] | tuple[float, float, float, float]
+
+It must return 4 values (percentages) corresponding to the experts:
+["Language", "Logic", "Social", "World"]
+
+Example return formats:
+- [12.5, 45.0, 22.5, 20.0]
+- {"Language": 12.5, "Logic": 45.0, "Social": 22.5, "World": 20.0}
+- (12.5, 45.0, 22.5, 20.0)
+"""
+import torch
+import numpy as np
+import torch.nn.functional as F
+from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig
+from typing import Union, Dict, List, Tuple
+
+from models.micro_olmo import MiCRoOLMo
+from models.micro_llama import MiCRoLlama
+from models.micro_moe_llama import MiCRoLlamaMoE
+
+DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'
+
+def get_expert_routing(model_id: str, hf_token: str, prompt: Union[str, List[Dict[str, str]]]) -> Union[List[float], Dict[str, float], Tuple[float, float, float, float]]:
+
+    model, tokenizer = build_model(model_id, hf_token)
+
+    if isinstance(prompt, str):
+        generation, routing_weights = generate_continuation(model, tokenizer, prompt)
+    elif isinstance(prompt, dict):
+        generation = None
+        routing_weights = get_routing_weights(model, tokenizer, [prompt])
+
+    model_routing_percentages = aggregate_routing_weights(routing_weights)
+
+    if generation is not None:
+        print(f"Generation:\n{generation}")
+    
+    return {
+        "Language": float(model_routing_percentages[3]),
+        "Logic": float(model_routing_percentages[0]),
+        "Social": float(model_routing_percentages[1]),
+        "World": float(model_routing_percentages[2]),
+    }, generation
+
+def get_model_path(model_name: str) -> Tuple[str, str, AutoModelForCausalLM]:
+    return {
+        # MiCRo-Llama
+        "micro-llama-1b": ("bkhmsi/micro-llama-1b", "meta-llama/Llama-3.2-1B-Instruct", MiCRoLlama),
+        "micro-llama-3b": ("bkhmsi/micro-llama-3b", "meta-llama/Llama-3.2-3B-Instruct", MiCRoLlama),
+        "micro-llama-1b-dpo": ("bkhmsi/micro-llama-1b-dpo", "meta-llama/Llama-3.2-1B-Instruct", MiCRoLlama),
+
+        # MiCRo-MoE-Llama
+        "micro-moe-llama-1b": ("bkhmsi/micro-moe-llama-1b", "meta-llama/Llama-3.2-1B-Instruct", MiCRoLlamaMoE),
+        
+        # MiCRo-OLMo
+        "micro-olmo": ("bkhmsi/micro-olmo-1b", "allenai/OLMo-2-0425-1B-Instruct", MiCRoOLMo),
+
+        # MiCRo-SmolLM2
+        "micro-smollm2-135m": ("bkhmsi/micro-smollm2-135m", "HuggingFaceTB/SmolLM2-135M-Instruct", MiCRoLlama),
+        "micro-smollm2-360m": ("bkhmsi/micro-smollm2-360m", "HuggingFaceTB/SmolLM2-360M-Instruct", MiCRoLlama),
+
+        # MiCRo-MoE-SmolLM2
+        "micro-moe-smollm2-135m": ("bkhmsi/micro-moe-smollm2-135m", "HuggingFaceTB/SmolLM2-135M-Instruct", MiCRoLlamaMoE),
+        "micro-moe-smollm2-360m": ("bkhmsi/micro-moe-smollm2-360m", "HuggingFaceTB/SmolLM2-360M-Instruct", MiCRoLlamaMoE),
+    }.get(model_name, (model_name, model_name, AutoModelForCausalLM))
+
+def aggregate_routing_weights(routing_weights):
+    experts = ["Logic", "Social", "World", "Language"]
+    expert_token_model = np.zeros((len(experts)), dtype=int)
+    expert_layer_token = np.zeros((routing_weights.shape[0], len(experts)), dtype=int)
+    num_layers = routing_weights.shape[0]
+
+    for layer_idx in range(num_layers):
+        for token_idx in range(len(routing_weights[layer_idx])):
+            expert_idx = routing_weights[layer_idx][token_idx].argmax()
+            if layer_idx >= 2 and layer_idx < num_layers - 2:
+                expert_token_model[expert_idx] += 1
+            expert_layer_token[layer_idx][expert_idx] += 1
+    return expert_token_model, expert_layer_token
+
+def generate_continuation(model, 
+    tokenizer, 
+    prompts, 
+    max_tokens=1024,
+    use_cache=True, 
+    return_routing_weights=True
+):
+
+    if isinstance(prompts, str):
+        prompts = [{"role": "user", "content": prompts}]
+
+    tokenizer.padding_side = "left"
+    inputs = tokenizer.apply_chat_template([
+        prompt for prompt in prompts
+    ], return_tensors="pt", padding=True, add_generation_prompt=True).to(DEVICE)
+
+    attention_mask = torch.ones_like(inputs)
+    attention_mask[inputs == tokenizer.pad_token_id] = 0
+
+    outputs = model.generate(
+        input_ids=inputs,
+        attention_mask=attention_mask, 
+        max_new_tokens=max_tokens,
+        use_cache=use_cache,
+        stop_strings=["</s>","<|eot_id|>", "<|im_start|>user"],
+        tokenizer=tokenizer,
+        pad_token_id=tokenizer.pad_token_id,
+        temperature=0,
+        top_p=1.0,
+        do_sample=False,
+    )
+    
+    if return_routing_weights:
+        attention_mask = torch.ones_like(outputs)
+        attention_mask[outputs == tokenizer.pad_token_id] = 0
+        model_output = model(input_ids=outputs, attention_mask=attention_mask)
+        torch.cuda.empty_cache()
+
+        routing_weights = model_output.routing_weights        
+        routing_weights = np.concatenate([
+            F.softmax(rw, dim=-1)[:, inputs.shape[1]:].detach().float().cpu().numpy() 
+            for rw in routing_weights
+        ])
+        
+    else:
+        routing_weights = None
+
+    inputs_text = tokenizer.batch_decode(inputs, skip_special_tokens=False)
+
+    generations = []
+    for i, output in enumerate(outputs):
+        decoded_output = tokenizer.decode(output, skip_special_tokens=False)
+        decoded_output = decoded_output.replace(inputs_text[i], "")
+        decoded_output = decoded_output.replace(tokenizer.pad_token, "").strip()
+        decoded_output = decoded_output.replace("<|end_of_text|>", "").strip()
+        decoded_output = decoded_output.replace("<|endoftext|>", "").strip()
+        decoded_output = decoded_output.replace("<|eot_id|>", "").strip()
+        decoded_output = decoded_output.replace("\n<|im_start|>user", "").strip()
+        generations.append(decoded_output)
+
+    return (generations, routing_weights) if return_routing_weights else generations
+
+def get_routing_weights(model, tokenizer, prompts, apply_chat_template=True):
+    """
+    Get routing weights for the given prompts using the model.
+    Args:
+        model: The MiCRoLlama or MiCRoOLMo model.
+        tokenizer: The tokenizer for the model.
+        prompts: A string or list of dictionaries containing the prompts.
+    Returns:
+        routing_weights: A list of routing weights for each layer.
+    """
+
+    tokenizer.padding_side = "left"
+    if apply_chat_template:
+        if isinstance(prompts, str):
+            prompts = [{"role": "user", "content": prompts}]
+
+        inputs = tokenizer.apply_chat_template([
+            prompt for prompt in prompts
+        ], return_tensors="pt", padding=True).to(DEVICE)
+
+        input_without_response = tokenizer.apply_chat_template([
+                prompt[:-1] for prompt in prompts
+            ], return_tensors="pt", padding=True,
+        ).to(DEVICE)
+    else:
+        inputs = tokenizer(prompts[0] + prompts[1], return_tensors="pt", padding=True).input_ids.to(DEVICE)
+        input_without_response = tokenizer(prompts[0], return_tensors="pt", padding=True).input_ids.to(DEVICE)
+
+    attention_mask = torch.ones_like(inputs)
+    attention_mask[inputs == tokenizer.pad_token_id] = 0
+
+    model_output = model(input_ids=inputs, attention_mask=attention_mask)
+
+    routing_weights = model_output.routing_weights   
+    routing_weights = np.stack([F.softmax(rw, dim=-1).detach().float().cpu().numpy() for rw in routing_weights], axis=0).squeeze()
+
+    offset = len(input_without_response[0])-1
+    routing_weights = routing_weights[:, offset:-1]
+
+    return routing_weights
+
+def build_model(model_id: str, hf_token: str, use_cache: bool = True):
+
+    model_path, base_model, model_class = get_model_path(model_id)
+
+    model_config = AutoConfig.from_pretrained(base_model, use_auth_token=hf_token)
+    model_config.config_path = f"configs/{model_id}.yml"
+
+    model_config.torch_dtype = torch.bfloat16
+    model_config.use_bfloat16 = True
+    model_config._attn_implementation = "flash_attention_2"
+    model_config.use_cache = use_cache
+    model_config.ablate = []
+
+    tokenizer = AutoTokenizer.from_pretrained(base_model, use_auth_token=hf_token)
+    tokenizer.padding_side = "left"
+
+    if "llama" in model_id:
+        tokenizer.pad_token_id = 128004
+    if "olmo" in model_id:
+        tokenizer.pad_token_id = 100277
+        tokenizer.add_special_tokens({'additional_special_tokens': ['<|assistant|>']})
+    elif "smollm2" in model_id:
+        tokenizer.pad_token_id = 2
+    else:
+        tokenizer.pad_token_id = 128004
+
+    if "olmo" in model_id:
+        model_config.vocab_size = len(tokenizer)
+
+    model = model_class.from_pretrained(model_path, config=model_config, low_cpu_mem_usage=True)
+
+    model.to(DEVICE)
+    model = model.bfloat16()
+    model.eval()
+    return model, tokenizer