Add files using upload-large-folder tool

README.md

@@ -0,0 +1,81 @@
+---
+library_name: vllm
+pipeline_tag: text-generation
+tags:
+  - laguna
+  - vllm
+  - fp8
+  - moe
+---
+
+# Laguna-M.1 (FP8 + FP8 KV-cache)
+
+Poolside Laguna-M, FP8-quantized weights (block scheme, 128×128) with FP8 KV-cache scales.
+Packaged as a self-contained HuggingFace repo so partners can `hf download` once and
+serve via the bundled vLLM overlay image without additional build steps.
+
+## Contents
+
+| Path                                          | Size   | What it is                                                 |
+|-----------------------------------------------|--------|------------------------------------------------------------|
+| `config.json`, `*.safetensors`, `tokenizer.*` | 214 GB | Laguna-M FP8 weights + FP8 KV-cache scales + tokenizer     |
+| `image/laguna-vllm-v0.19.0-overlay.tar.gz`    | 8.9 GB | vLLM v0.19.0 Docker image with Laguna support overlaid     |
+
+## Hardware
+
+Built and tested on **NVIDIA H200 (141 GB HBM3e)**. Recommended serving config: **4× H200, TP=4**.
+Other GPU generations (H100, B200) may work but are untested.
+
+## Quickstart
+
+```bash
+# 1. Download.
+hf download poolside/Laguna-M.1 --local-dir laguna-m
+
+# 2. Load the vLLM image.
+gunzip -c laguna-m/image/laguna-vllm-v0.19.0-overlay.tar.gz | docker load
+# -> laguna-vllm:v0.19.0-overlay
+
+# 3. Serve on 4× H200.
+docker run --gpus '"device=0,1,2,3"' --network host \
+  -v "$PWD/laguna-m":/model \
+  laguna-vllm:v0.19.0-overlay \
+  /model \
+  --tensor-parallel-size 4 \
+  --trust-remote-code \
+  --dtype bfloat16 \
+  --kv-cache-dtype fp8 \
+  --max-model-len 4096 \
+  --gpu-memory-utilization 0.85 \
+  --served-model-name laguna \
+  --reasoning-parser poolside_v1 \
+  --tool-call-parser poolside_v1 \
+  --enable-auto-tool-choice \
+  --port 8000
+```
+
+Smoke test once the server logs `Application startup complete`:
+
+```bash
+curl -s http://127.0.0.1:8000/v1/completions \
+  -H 'Content-Type: application/json' \
+  -d '{"model":"laguna","prompt":"The capital of France is","max_tokens":8,"temperature":0}'
+```
+
+## Eval (5-shot)
+
+- MMLU: 79.06% ± 0.33%  *(all 14k questions, TP=4 H200)*
+- GSM8K: 92% strict-match  *(n=50 smoke, TP=4 H200; full run to come)*
+
+## Architecture
+
+Laguna-M is a 70-layer MoE transformer:
+
+- hidden=4096, Q-heads=64, KV-heads=8, head_dim=128 (Q-projection is 2× hidden → 8192)
+- First 3 layers dense SwiGLU; remaining 67 are sparse MoE (256 experts, top-k=16, sigmoid router, shared expert)
+- Per-element attention output gating (softplus)
+- RoPE θ=500000, YaRN factor=32, original ctx 4096
+- Auxiliary-loss-free load balancing via `e_score_correction_bias`
+
+Weights use compressed-tensors FP8 block format (128×128 blocks, dynamic activation quant).
+FP8 KV-cache scales are included under `model.layers.*.self_attn.{k,v}_scale`.

image/laguna-vllm-v0.19.0-overlay.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3a3dd628330235925fa2d9bb07c7438c41236ffc6f18e6e079f1bd0a4e019814
+size 9548969617

modeling_laguna.py

@@ -0,0 +1,671 @@
+# ruff: noqa
+# Copyright 2025 Poolside and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Laguna model implementation for transformers 4.56-4.x (used by vLLM).
+
+This avoids v5-only APIs (use_kernel_forward_from_hub, create_causal_mask,
+dynamic_rope_update, auto_docstring, can_return_tuple, etc.) while keeping
+the architecture identical to the v5 version.
+"""
+
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+from transformers.generation import GenerationMixin
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.utils.generic import OutputRecorder, check_model_inputs
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.modeling_outputs import MoeModelOutputWithPast, MoeCausalLMOutputWithPast
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
+
+from .configuration_laguna import LagunaConfig
+
+
+class LagunaRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        LagunaRMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+class LagunaRotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, config: LagunaConfig, device=None):
+        super().__init__()
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+
+        # v4 uses rope_theta + rope_scaling (top-level config fields)
+        rope_type = "default"
+        if config.rope_scaling is not None:
+            rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type", "default"))
+
+        self.rope_type = rope_type
+        if self.rope_type == "default":
+            inv_freq, self.attention_scaling = self._compute_default_rope_parameters(config, device)
+        else:
+            rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+            inv_freq, self.attention_scaling = rope_init_fn(config, device)
+
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
+
+    @staticmethod
+    def _compute_default_rope_parameters(
+        config: LagunaConfig,
+        device: Optional["torch.device"] = None,
+    ) -> tuple["torch.Tensor", float]:
+        base = config.rope_theta
+        dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
+        attention_factor = 1.0
+        inv_freq = 1.0 / (
+            base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim)
+        )
+        return inv_freq, attention_factor
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+class LagunaMLP(nn.Module):
+    def __init__(self, config, intermediate_size=None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size if intermediate_size is None else intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+class LagunaTopKRouter(nn.Module):
+    """Laguna MoE router using sigmoid scoring (not softmax)."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.top_k = config.num_experts_per_tok
+        self.num_experts = config.num_experts
+        self.norm_topk_prob = config.norm_topk_prob
+        self.hidden_dim = config.hidden_size
+        self.weight = nn.Parameter(torch.zeros(self.num_experts, self.hidden_dim))
+
+    def forward(self, hidden_states: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        hidden_states = hidden_states.reshape(-1, self.hidden_dim)
+        router_logits = F.linear(hidden_states, self.weight)
+        # Laguna-specific: sigmoid routing in float32 for precision
+        routing_weights = torch.sigmoid(router_logits.float())
+        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+        if self.norm_topk_prob:
+            routing_weights = routing_weights / routing_weights.sum(dim=-1, keepdim=True)
+        routing_weights = routing_weights.to(hidden_states.dtype)
+        return router_logits, routing_weights, selected_experts
+
+
+class LagunaSparseMoeBlock(nn.Module):
+    """Laguna MoE block using sigmoid router, per-expert MLPs, and a shared expert."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.num_experts = config.num_experts
+        self.top_k = config.num_experts_per_tok
+        self.gate = LagunaTopKRouter(config)
+        self.experts = nn.ModuleList(
+            [LagunaMLP(config, intermediate_size=config.moe_intermediate_size) for _ in range(self.num_experts)]
+        )
+        self.shared_expert = LagunaMLP(config, intermediate_size=config.shared_expert_intermediate_size)
+        self.shared_expert_gate = (
+            nn.Linear(config.hidden_size, 1, bias=False) if getattr(config, "moe_shared_gate", False) else None
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+
+        shared_expert_output = self.shared_expert(hidden_states)
+        if self.shared_expert_gate is not None:
+            shared_expert_output = shared_expert_output * torch.sigmoid(self.shared_expert_gate(hidden_states))
+
+        # Routed experts
+        _, routing_weights, selected_experts = self.gate(hidden_states)
+        final_hidden_states = torch.zeros_like(hidden_states)
+
+        expert_mask = F.one_hot(selected_experts, num_classes=self.num_experts)
+        expert_mask = expert_mask.permute(2, 1, 0)
+
+        for expert_idx in range(self.num_experts):
+            top_k_pos, token_idx = torch.where(expert_mask[expert_idx])
+            if token_idx.shape[0] == 0:
+                continue
+            current_state = hidden_states[token_idx]
+            current_hidden_states = self.experts[expert_idx](current_state)
+            current_hidden_states = current_hidden_states * routing_weights[token_idx, top_k_pos, None]
+            final_hidden_states.index_add_(0, token_idx, current_hidden_states.to(final_hidden_states.dtype))
+
+        final_hidden_states = final_hidden_states + shared_expert_output
+        final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
+        return final_hidden_states
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors."""
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: torch.Tensor | None,
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs,
+):
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+class LagunaAttention(nn.Module):
+    """Laguna attention with QK normalization and output gating."""
+
+    def __init__(self, config: LagunaConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.head_dim = config.head_dim
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+        self.is_causal = True
+
+        # Laguna: no QKV bias, explicit head_dim
+        self.q_proj = nn.Linear(config.hidden_size, config.num_attention_heads * config.head_dim, bias=False)
+        self.k_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * config.head_dim, bias=False)
+        self.v_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * config.head_dim, bias=False)
+        self.o_proj = nn.Linear(config.num_attention_heads * config.head_dim, config.hidden_size, bias=False)
+        # Laguna-specific: gating projection
+        self.g_proj = nn.Linear(config.hidden_size, config.num_attention_heads * config.head_dim, bias=False)
+        # QK normalization (RMSNorm applied per-head after reshape, before RoPE)
+        self.q_norm = LagunaRMSNorm(config.head_dim, eps=config.rms_norm_eps)
+        self.k_norm = LagunaRMSNorm(config.head_dim, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: torch.Tensor | None,
+        past_key_values: Cache | None = None,
+        cache_position: torch.LongTensor | None = None,
+        **kwargs,
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(hidden_shape).transpose(1, 2)
+        key_states = key_states.view(hidden_shape).transpose(1, 2)
+        value_states = value_states.view(hidden_shape).transpose(1, 2)
+
+        # QK normalization (applied per-head before RoPE)
+        query_states = self.q_norm(query_states)
+        key_states = self.k_norm(key_states)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_values is not None:
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        attention_interface = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+
+        # Laguna-specific: apply gating BEFORE o_proj
+        gate = F.softplus(self.g_proj(hidden_states).float()).to(attn_output.dtype)
+        attn_output = attn_output * gate
+
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, attn_weights
+
+
+class LagunaDecoderLayer(nn.Module):
+    """Laguna decoder layer with gated attention and sigmoid-routed MoE."""
+
+    def __init__(self, config: LagunaConfig, layer_idx: int):
+        super().__init__()
+        self.self_attn = LagunaAttention(config, layer_idx)
+        # Use MoE or dense MLP based on layer configuration
+        if (layer_idx not in config.mlp_only_layers) and (
+            config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0
+        ):
+            self.mlp = LagunaSparseMoeBlock(config)
+        else:
+            self.mlp = LagunaMLP(config, intermediate_size=config.intermediate_size)
+        self.input_layernorm = LagunaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = LagunaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.hidden_size = config.hidden_size
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        use_cache: bool | None = False,
+        cache_position: torch.LongTensor | None = None,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        # Self Attention
+        hidden_states, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+
+class LagunaPreTrainedModel(PreTrainedModel):
+    config_class = LagunaConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["LagunaDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_cache_class = True
+    _can_record_outputs = {
+        "router_logits": OutputRecorder(LagunaTopKRouter, index=0),
+        "hidden_states": LagunaDecoderLayer,
+        "attentions": LagunaAttention,
+    }
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, LagunaTopKRouter):
+            module.weight.data.normal_(mean=0.0, std=std)
+
+
+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,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """Create 4D causal mask from 2D attention mask, compatible with transformers 4.x."""
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # Already a 4D mask
+        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=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()
+            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 LagunaModel(LagunaPreTrainedModel):
+    def __init__(self, config: LagunaConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [LagunaDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = LagunaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = LagunaRotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @check_model_inputs
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        use_cache: bool | None = None,
+        cache_position: torch.LongTensor | None = None,
+        **kwargs,
+    ):
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        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 = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=inputs_embeds.shape[1],
+            target_length=cache_position[-1].item() + 1 if cache_position is not None else inputs_embeds.shape[1],
+            dtype=inputs_embeds.dtype,
+            device=inputs_embeds.device,
+            cache_position=cache_position,
+            batch_size=inputs_embeds.shape[0],
+        )
+
+        hidden_states = inputs_embeds
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+            if self.gradient_checkpointing and self.training:
+                hidden_states = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    causal_mask,
+                    position_ids,
+                    past_key_values,
+                    use_cache,
+                    cache_position,
+                    position_embeddings,
+                )
+            else:
+                hidden_states = decoder_layer(
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    past_key_values=past_key_values,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                    position_embeddings=position_embeddings,
+                    **kwargs,
+                )
+
+        hidden_states = self.norm(hidden_states)
+
+        return MoeModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+        )
+
+
+def load_balancing_loss_func(
+    gate_logits: torch.Tensor | tuple[torch.Tensor] | None,
+    num_experts: int | None = None,
+    top_k=2,
+    attention_mask: torch.Tensor | None = None,
+) -> torch.Tensor | int:
+    r"""
+    Computes auxiliary load balancing loss as in Switch Transformer.
+
+    See Switch Transformer (https://huggingface.co/papers/2101.03961) for more details.
+    """
+    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:
+        tokens_per_expert = torch.mean(expert_mask.float(), dim=0)
+        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)
+
+        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)
+        )
+
+        tokens_per_expert = torch.sum(expert_mask.float() * expert_attention_mask, dim=0) / torch.sum(
+            expert_attention_mask, dim=0
+        )
+
+        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)
+        )
+
+        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
+
+
+class LagunaForCausalLM(LagunaPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
+    _tp_plan = {"lm_head": "colwise_rep"}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = LagunaModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.router_aux_loss_coef = config.router_aux_loss_coef
+        self.num_experts = config.num_experts
+        self.num_experts_per_tok = config.num_experts_per_tok
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        labels: torch.LongTensor | None = None,
+        use_cache: bool | None = None,
+        output_router_logits: bool | None = None,
+        cache_position: torch.LongTensor | None = None,
+        logits_to_keep: int | torch.Tensor = 0,
+        **kwargs,
+    ) -> MoeCausalLMOutputWithPast:
+        r"""
+        Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+        config.vocab_size]` or -100. Tokens with indices set to `-100` are ignored (masked), the loss is
+        only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+        """
+        output_router_logits = (
+            output_router_logits if output_router_logits is not None else self.config.output_router_logits
+        )
+
+        outputs: MoeModelOutputWithPast = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_router_logits=output_router_logits,
+            cache_position=cache_position,
+            **kwargs,
+        )
+
+        hidden_states = outputs.last_hidden_state
+        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, labels, self.vocab_size, **kwargs)
+
+        aux_loss = None
+        if output_router_logits:
+            aux_loss = load_balancing_loss_func(
+                outputs.router_logits,
+                self.num_experts,
+                self.num_experts_per_tok,
+                attention_mask,
+            )
+            if labels is not None and isinstance(aux_loss, torch.Tensor):
+                loss += self.router_aux_loss_coef * aux_loss.to(loss.device)
+
+        return MoeCausalLMOutputWithPast(
+            loss=loss,
+            aux_loss=aux_loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            router_logits=outputs.router_logits,
+        )
+
+
+__all__ = ["LagunaForCausalLM", "LagunaModel", "LagunaPreTrainedModel"]

serve.sh

@@ -0,0 +1,26 @@
+#!/usr/bin/env bash
+# Minimal serve command for Laguna-M.1 on 4× H200.
+# Expects the bundled image to be loaded (`docker load < image/...tar.gz`)
+# and this script run from the downloaded repo root.
+set -euo pipefail
+
+MODEL_DIR="${MODEL_DIR:-$(cd "$(dirname "$0")" && pwd)}"
+PORT="${PORT:-8000}"
+GPUS="${GPUS:-0,1,2,3}"
+
+docker run --rm --gpus "\"device=${GPUS}\"" --network host \
+  -v "${MODEL_DIR}":/model \
+  laguna-vllm:v0.19.0-overlay \
+  /model \
+  --tensor-parallel-size 4 \
+  --trust-remote-code \
+  --dtype bfloat16 \
+  --kv-cache-dtype fp8 \
+  --max-model-len 4096 \
+  --gpu-memory-utilization 0.85 \
+  --served-model-name laguna \
+  --reasoning-parser poolside_v1 \
+  --tool-call-parser poolside_v1 \
+  --enable-auto-tool-choice \
+  --default-chat-template-kwargs '{"enable_thinking": true}' \
+  --port "${PORT}"

special_tokens_map.json

@@ -0,0 +1,9 @@
+{
+  "bos_token": "〈|EOS|〉",
+  "cls_token": "〈|CLS|〉",
+  "eos_token": "〈|EOS|〉",
+  "mask_token": "〈|MASK|〉",
+  "pad_token": "〈|PAD|〉",
+  "sep_token": "〈|SEP|〉",
+  "unk_token": "〈|UNK|〉"
+}

tokenizer_config.json

@@ -0,0 +1,576 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "〈|UNK|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "〈|CODE_START|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "〈|EOS|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "3": {
+      "content": "〈|CODE_END|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "4": {
+      "content": "〈|META_START|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "5": {
+      "content": "〈|META_END|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "6": {
+      "content": "〈|FIM_MIDDLE|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "7": {
+      "content": "〈|FIM_SUFFIX|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "8": {
+      "content": "〈|SEP|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "9": {
+      "content": "〈|PAD|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "10": {
+      "content": "〈|CLS|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "11": {
+      "content": "〈|FIM_START|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "12": {
+      "content": "〈|MASK|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "13": {
+      "content": "|◊|",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "14": {
+      "content": "〈|",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "15": {
+      "content": "|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "16": {
+      "content": "〈|/",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "17": {
+      "content": "/|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "20": {
+      "content": "〈|SPECIAL_1|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "21": {
+      "content": "〈|SPECIAL_2|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "22": {
+      "content": "〈|SPECIAL_3|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "27": {
+      "content": "〈|SPECIAL_8|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "28": {
+      "content": "〈|SPECIAL_9|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "29": {
+      "content": "〈|SPECIAL_10|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "30": {
+      "content": "〈|SPECIAL_11|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "31": {
+      "content": "〈|SPECIAL_12|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32": {
+      "content": "〈|SPECIAL_13|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "33": {
+      "content": "〈|SPECIAL_14|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "34": {
+      "content": "〈|SPECIAL_15|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "35": {
+      "content": "〈|SPECIAL_16|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "36": {
+      "content": "〈|SPECIAL_17|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "37": {
+      "content": "〈|SPECIAL_18|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "38": {
+      "content": "〈|SPECIAL_19|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "39": {
+      "content": "〈|SPECIAL_20|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "40": {
+      "content": "〈|SPECIAL_21|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "41": {
+      "content": "〈|SPECIAL_22|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "42": {
+      "content": "〈|SPECIAL_23|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "43": {
+      "content": "〈|SPECIAL_24|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "44": {
+      "content": "〈|SPECIAL_25|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "45": {
+      "content": "〈|SPECIAL_26|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "46": {
+      "content": "〈|SPECIAL_27|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "47": {
+      "content": "〈|SPECIAL_28|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "48": {
+      "content": "〈|SPECIAL_29|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49": {
+      "content": "〈|SPECIAL_30|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50": {
+      "content": "〈|SPECIAL_31|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "51": {
+      "content": "〈|SPECIAL_32|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "52": {
+      "content": "〈|SPECIAL_33|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "53": {
+      "content": "〈|SPECIAL_34|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "54": {
+      "content": "〈|SPECIAL_35|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "55": {
+      "content": "〈|SPECIAL_36|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "56": {
+      "content": "〈|SPECIAL_37|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "57": {
+      "content": "〈|SPECIAL_38|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "58": {
+      "content": "〈|SPECIAL_39|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "59": {
+      "content": "〈|SPECIAL_40|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "60": {
+      "content": "〈|SPECIAL_41|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "61": {
+      "content": "〈|SPECIAL_42|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "62": {
+      "content": "〈|SPECIAL_43|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "63": {
+      "content": "〈|SPECIAL_44|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "64": {
+      "content": "〈|SPECIAL_45|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65": {
+      "content": "〈|SPECIAL_46|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "66": {
+      "content": "〈|SPECIAL_47|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "67": {
+      "content": "〈|SPECIAL_48|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "68": {
+      "content": "〈|SPECIAL_49|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "69": {
+      "content": "〈|SPECIAL_50|〉",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "18": {
+      "content": "<think>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "19": {
+      "content": "</think>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "23": {
+      "content": "<assistant>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "24": {
+      "content": "</assistant>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "25": {
+      "content": "<tool_call>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "26": {
+      "content": "</tool_call>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    }
+  },
+  "bos_token": "〈|EOS|〉",
+  "clean_up_tokenization_spaces": false,
+  "cls_token": "〈|CLS|〉",
+  "eos_token": "〈|EOS|〉",
+  "extra_special_tokens": {},
+  "mask_token": "〈|MASK|〉",
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "〈|PAD|〉",
+  "sep_token": "〈|SEP|〉",
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "unk_token": "〈|UNK|〉",
+  "chat_template": "{% include 'chat_template.jinja' %}"
+}