Upload checkpoint on step 0177557

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Michael Lee (李登淳)
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

chat_template.jinja

@@ -0,0 +1,15 @@
+{% for message in messages %}
+{% if message['role'] == 'user' %}
+{{ '<|user|>
+' + message['content'] + eos_token }}
+{% elif message['role'] == 'system' %}
+{{ '<|system|>
+' + message['content'] + eos_token }}
+{% elif message['role'] == 'assistant' %}
+{{ '<|assistant|>
+'  + message['content'] + eos_token }}
+{% endif %}
+{% if loop.last and add_generation_prompt %}
+{{ '<|assistant|>' }}
+{% endif %}
+{% endfor %}

config.json

@@ -0,0 +1,28 @@
+{
+  "architectures": [
+    "TinyMixtralForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "dtype": "float32",
+  "expert_intermediate_size": 2389,
+  "head_dim": 64,
+  "hidden_size": 896,
+  "initializer_range": 0.02,
+  "max_position_embeddings": 2048,
+  "model_type": "tinymixtral",
+  "num_attention_heads": 14,
+  "num_experts_per_tok": 2,
+  "num_hidden_layers": 10,
+  "num_key_value_heads": 2,
+  "num_local_experts": 6,
+  "rms_norm_eps": 1e-06,
+  "rope_theta": 1000000.0,
+  "router_aux_loss_coef": 0.01,
+  "router_jitter_noise": 0.01,
+  "transformers_version": "4.57.3",
+  "vocab_size": 32000,
+  "auto_map": {
+    "AutoConfig": "configuration_tinymixtral.TinyMixtralConfig",
+    "AutoModelForCausalLM": "modeling_tinymixtral.TinyMixtralForCausalLM"
+  }
+}

configuration_tinymixtral.py

@@ -0,0 +1,47 @@
+# Copyright (C) Michael Lee (李登淳) 2026. All rights reserved.
+# Open-source under the MIT License. See LICENSE for details.
+
+from transformers import PretrainedConfig
+
+
+class TinyMixtralConfig(PretrainedConfig):
+    model_type = "tinymixtral"
+
+    def __init__(
+        self,
+        vocab_size: int = 32000,
+        hidden_size: int = 896,
+        num_hidden_layers: int = 10,
+        num_attention_heads: int = 14,
+        num_key_value_heads: int = 2,
+        head_dim: int = 64,
+        max_position_embeddings: int = 2048,
+        num_local_experts: int = 6,
+        num_experts_per_tok: int = 2,
+        expert_intermediate_size: int = 2389,
+        router_aux_loss_coef: float = 0.01,
+        router_jitter_noise: float = 0.01,
+        rms_norm_eps: float = 1e-6,
+        rope_theta: float = 1_000_000.0,
+        attention_dropout: float = 0.0,
+        tie_word_embeddings: bool = True,
+        initializer_range: float = 0.02,
+        **kwargs,
+    ):
+        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.head_dim = head_dim
+        self.max_position_embeddings = max_position_embeddings
+        self.num_local_experts = num_local_experts
+        self.num_experts_per_tok = num_experts_per_tok
+        self.expert_intermediate_size = expert_intermediate_size
+        self.router_aux_loss_coef = router_aux_loss_coef
+        self.router_jitter_noise = router_jitter_noise
+        self.rms_norm_eps = rms_norm_eps
+        self.rope_theta = rope_theta
+        self.attention_dropout = attention_dropout
+        self.initializer_range = initializer_range

modeling_tinymixtral.py

@@ -0,0 +1,237 @@
+# Copyright (C) Michael Lee (李登淳) 2026. All rights reserved.
+# Open-source under the MIT License. See LICENSE for details.
+
+from dataclasses import dataclass
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.utils.checkpoint import checkpoint
+from transformers import PreTrainedModel
+
+from .configuration_tinymixtral import TinyMixtralConfig
+
+
+# ============================================================
+# Layers
+# ============================================================
+
+class RMSNorm(nn.Module):
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(dim))
+        self.eps = eps
+
+    def forward(self, x):
+        dtype = x.dtype
+        x = x.float()
+        norm = x.pow(2).mean(-1, keepdim=True)
+        x = x * torch.rsqrt(norm + self.eps)
+        return (x * self.weight).to(dtype)
+
+
+class RotaryEmbedding(nn.Module):
+    def __init__(self, dim, max_position_embeddings=2048, theta=10000.0):
+        super().__init__()
+        self.dim = dim
+        self.max_position_embeddings = max_position_embeddings
+        self.theta = theta
+        self._build_cache()
+
+    def _build_cache(self):
+        inv_freq = 1.0 / (self.theta ** (torch.arange(0, self.dim, 2).float() / self.dim))
+        t = torch.arange(self.max_position_embeddings).float()
+        freqs = torch.outer(t, inv_freq)
+        emb = torch.cat((freqs, freqs), dim=-1)
+        self.register_buffer("cos_cached", emb.cos(), persistent=False)
+        self.register_buffer("sin_cached", emb.sin(), persistent=False)
+
+    def forward(self, x, position_ids):
+        cos = self.cos_cached[position_ids].unsqueeze(1)
+        sin = self.sin_cached[position_ids].unsqueeze(1)
+        x_rot = x.float()
+        x1, x2 = x_rot.chunk(2, dim=-1)
+        rotated = torch.cat((-x2, x1), dim=-1)
+        return (x_rot * cos + rotated * sin).to(x.dtype)
+
+
+class GQAAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.num_kv_heads = config.num_key_value_heads
+        self.head_dim = config.head_dim
+        self.num_groups = self.num_heads // self.num_kv_heads
+        assert self.num_heads % self.num_kv_heads == 0
+
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+        self.rotary_emb = RotaryEmbedding(self.head_dim, config.max_position_embeddings, config.rope_theta)
+        self.attention_dropout = config.attention_dropout
+
+    def forward(self, hidden_states, attention_mask=None, position_ids=None):
+        B, S, _ = hidden_states.shape
+        q = self.q_proj(hidden_states).view(B, S, self.num_heads, self.head_dim).transpose(1, 2)
+        k = self.k_proj(hidden_states).view(B, S, self.num_kv_heads, self.head_dim).transpose(1, 2)
+        v = self.v_proj(hidden_states).view(B, S, self.num_kv_heads, self.head_dim).transpose(1, 2)
+        k = k.unsqueeze(2).expand(-1, -1, self.num_groups, -1, -1).reshape(B, self.num_heads, S, self.head_dim)
+        v = v.unsqueeze(2).expand(-1, -1, self.num_groups, -1, -1).reshape(B, self.num_heads, S, self.head_dim)
+        if position_ids is None:
+            position_ids = torch.arange(S, device=hidden_states.device).unsqueeze(0).expand(B, -1)
+        q, k = self.rotary_emb(q, position_ids), self.rotary_emb(k, position_ids)
+
+        if attention_mask is not None:
+            causal = torch.tril(torch.ones(S, S, device=hidden_states.device, dtype=torch.bool))
+            combined = causal[None, None, :, :] & attention_mask[:, None, None, :]
+            attn = F.scaled_dot_product_attention(
+                q, k, v, attn_mask=combined,
+                dropout_p=self.attention_dropout if self.training else 0.0,
+                is_causal=False,
+            )
+        else:
+            attn = F.scaled_dot_product_attention(
+                q, k, v, attn_mask=None,
+                dropout_p=self.attention_dropout if self.training else 0.0,
+                is_causal=True,
+            )
+        return self.o_proj(attn.transpose(1, 2).reshape(B, S, -1))
+
+
+class SparseMoE(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.num_experts = config.num_local_experts
+        self.top_k = config.num_experts_per_tok
+        self.expert_intermediate = config.expert_intermediate_size
+        self.jitter_noise = config.router_jitter_noise
+        self.aux_loss_coef = config.router_aux_loss_coef
+        self.router = nn.Linear(self.hidden_size, self.num_experts, bias=False)
+        self.gate_proj = nn.Parameter(torch.empty(self.num_experts, self.expert_intermediate, self.hidden_size))
+        self.up_proj = nn.Parameter(torch.empty(self.num_experts, self.expert_intermediate, self.hidden_size))
+        self.down_proj = nn.Parameter(torch.empty(self.num_experts, self.hidden_size, self.expert_intermediate))
+        self._init_weights()
+
+    def _init_weights(self, std=0.02):
+        nn.init.normal_(self.gate_proj, std=std)
+        nn.init.normal_(self.up_proj, std=std)
+        nn.init.normal_(self.down_proj, std=std)
+
+    def forward(self, x):
+        B, S, D = x.shape
+        x_flat = x.view(-1, D)
+        logits = self.router(x_flat)
+        if self.training and self.jitter_noise > 0:
+            logits = logits * (1 + torch.randn_like(logits) * self.jitter_noise)
+        weights = F.softmax(logits.float(), dim=-1).to(x.dtype)
+        w_topk, experts = torch.topk(weights, self.top_k, dim=-1)
+        w_topk = w_topk / w_topk.sum(dim=-1, keepdim=True)
+
+        aux = torch.tensor(0.0, device=x.device, dtype=x.dtype)
+        if self.training and self.aux_loss_coef > 0:
+            with torch.no_grad():
+                mask = F.one_hot(experts, num_classes=self.num_experts).float()
+                f_i = mask.mean(dim=(0, 1))
+            P_i = weights.mean(dim=0)
+            aux = (f_i.detach() * P_i).sum() * self.num_experts
+
+        out = torch.zeros(B * S, D, device=x.device, dtype=x.dtype)
+        for k in range(self.top_k):
+            for e in range(self.num_experts):
+                m = (experts[:, k] == e)
+                if not m.any():
+                    continue
+                ts = x_flat[m]
+                gate = F.silu(ts @ self.gate_proj[e].T)
+                up = ts @ self.up_proj[e].T
+                out[m] += (gate * up @ self.down_proj[e].T) * w_topk[m, k].unsqueeze(-1)
+        return out.view(B, S, D), aux
+
+
+class MoETransformerBlock(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.input_layernorm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.self_attn = GQAAttention(config)
+        self.moe = SparseMoE(config)
+
+    def forward(self, x, attention_mask=None, position_ids=None):
+        x = x + self.self_attn(self.input_layernorm(x), attention_mask, position_ids)
+        h, aux = self.moe(self.post_attention_layernorm(x))
+        return x + h, aux
+
+
+# ============================================================
+# Causal LM
+# ============================================================
+
+@dataclass
+class CausalLMOutputWithPast:
+    loss: Optional[torch.Tensor] = None
+    logits: torch.Tensor = None
+
+
+class TinyMixtralForCausalLM(PreTrainedModel):
+    config_class = TinyMixtralConfig
+    base_model_prefix = "tinymixtral"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["MoETransformerBlock"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = nn.ModuleList([MoETransformerBlock(config) for _ in range(config.num_hidden_layers)])
+        self.norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        if config.tie_word_embeddings:
+            self.lm_head.weight = self.embed_tokens.weight
+        self._use_activation_checkpointing = False
+        self.post_init()
+
+    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)
+
+    def gradient_checkpointing_enable(self, gradient_checkpointing_kwargs=None):
+        self._use_activation_checkpointing = True
+
+    def gradient_checkpointing_disable(self):
+        self._use_activation_checkpointing = False
+
+    def forward(self, input_ids, attention_mask=None, labels=None, return_dict=True, **kwargs):
+        B, S = input_ids.shape
+        pos = torch.arange(S, device=input_ids.device).unsqueeze(0).expand(B, -1)
+        cmask = attention_mask.bool() if attention_mask is not None else None
+
+        h = self.embed_tokens(input_ids)
+        total_aux = torch.tensor(0.0, device=input_ids.device, dtype=torch.float32)
+        for layer in self.layers:
+            if self._use_activation_checkpointing and self.training:
+                h, aux = checkpoint(layer, h, cmask, pos, use_reentrant=False)
+            else:
+                h, aux = layer(h, cmask, pos)
+            total_aux = total_aux + aux
+        logits = self.lm_head(self.norm(h)).float()
+
+        loss = None
+        if labels is not None:
+            loss = F.cross_entropy(
+                logits.reshape(-1, logits.size(-1)),
+                labels.reshape(-1),
+                ignore_index=-100,
+            )
+            loss = loss + self.config.router_aux_loss_coef * total_aux
+
+        if not return_dict:
+            return (loss, logits) if loss is not None else (logits,)
+        return CausalLMOutputWithPast(loss=loss, logits=logits)

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad3e467d0a029b27fc6e13c44d1e36ac1552cc183996aae04161215c47876864
+size 1729603131

special_tokens_map.json

@@ -0,0 +1,30 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347
+size 499723

tokenizer_config.json

@@ -0,0 +1,43 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "add_prefix_space": null,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "extra_special_tokens": {},
+  "legacy": false,
+  "model_max_length": 2048,
+  "pad_token": "</s>",
+  "padding_side": "right",
+  "sp_model_kwargs": {},
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": "<unk>",
+  "use_default_system_prompt": false
+}