Fix: vocab_size=32768 from tokenizer; top-level weights; ZeroDivisionError guard; all_tied_weights_keys->dict

config.json

@@ -9,7 +9,6 @@
   },
   "n_layers": 14,
   "d_model": 768,
-  "vocab_size": 32768,
   "attention_n_heads": 12,
   "attention_n_kv_heads": 1,
   "max_seq_len": 512,
@@ -19,5 +18,6 @@
   "norm_eps": 1e-05,
   "dropout": 0.1,
   "torch_dtype": "float32",
-  "transformers_version": "4.48.3"
+  "transformers_version": "4.48.3",
+  "vocab_size": 32768
 }

pico_decoder.py

@@ -1,11 +1,17 @@
+# pico_decoder.py — BeetleLM HuggingFace wrapper
+# Source: pico-lm/pico-train (Apache 2.0)
+# Load: AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True)
+
 from dataclasses import asdict
 from typing import TYPE_CHECKING, Any, Dict, Optional, Tuple, Union
+
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
 from torch.nn.attention import SDPBackend, sdpa_kernel
 from transformers import PretrainedConfig, PreTrainedModel
 from transformers.modeling_outputs import CausalLMOutput, CausalLMOutputWithPast
+
 try:
     if TYPE_CHECKING:
         from src.config import ModelConfig
@@ -13,43 +19,63 @@ except ImportError:
     pass
 
 
+# ── RMSNorm ──────────────────────────────────────────────────────────────────
+
 class RMSNorm(torch.nn.Module):
     def __init__(self, config):
         super().__init__()
-        self.eps = config.norm_eps
+        self.eps    = config.norm_eps
         self.weight = nn.Parameter(torch.ones(config.d_model))
+
     def _norm(self, x):
         return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
     def forward(self, x):
         return self._norm(x.float()).type_as(x) * self.weight
 
 
+# ── RoPE ─────────────────────────────────────────────────────────────────────
+
 class RoPE(nn.Module):
     _freqs_cis_tensor = None
+
     def __init__(self, config):
         super().__init__()
         self.theta = config.position_emb_theta
         self.dim   = config.d_model // config.attention_n_heads
         if RoPE._freqs_cis_tensor is None:
-            RoPE._freqs_cis_tensor = self._setup_freqs_cis(config.max_seq_len, self.theta, self.dim)
+            RoPE._freqs_cis_tensor = self._setup_freqs_cis(
+                config.max_seq_len, self.theta, self.dim
+            )
         self.register_buffer("_freqs_cis", RoPE._freqs_cis_tensor, persistent=False)
+
     @classmethod
     def _setup_freqs_cis(cls, seq_len, theta, dim):
-        _freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
-        return torch.polar(torch.ones_like(f := torch.outer(torch.arange(seq_len), _freqs)), f)
+        _freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: dim // 2].float() / dim))
+        freqs  = torch.outer(torch.arange(seq_len), _freqs)
+        return torch.polar(torch.ones_like(freqs), freqs)
+
     def get_freqs_cis(self, input_shape, start_pos, end_pos):
-        _f = self._freqs_cis[start_pos:end_pos]
+        _f   = self._freqs_cis[start_pos:end_pos]
         ndim = len(input_shape)
-        assert 0 <= 1 < ndim and _f.shape == (input_shape[1], input_shape[-1])
-        return _f.view(*[d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(input_shape)])
+        assert 0 <= 1 < ndim
+        assert _f.shape == (input_shape[1], input_shape[-1])
+        shape = [d if i == 1 or i == ndim - 1 else 1
+                 for i, d in enumerate(input_shape)]
+        return _f.view(*shape)
+
     def forward(self, queries, keys, start_pos=0):
         q_ = torch.view_as_complex(queries.float().reshape(*queries.shape[:-1], -1, 2))
         k_ = torch.view_as_complex(keys.float().reshape(*keys.shape[:-1], -1, 2))
-        fc = self.get_freqs_cis(q_.shape, start_pos, start_pos + q_.shape[1])
-        return (torch.view_as_real(q_ * fc).flatten(3).type_as(queries),
-                torch.view_as_real(k_ * fc).flatten(3).type_as(keys))
+        input_shape = q_.shape
+        fc = self.get_freqs_cis(input_shape, start_pos, start_pos + q_.shape[1])
+        q_rot = torch.view_as_real(q_ * fc).flatten(3)
+        k_rot = torch.view_as_real(k_ * fc).flatten(3)
+        return q_rot.type_as(queries), k_rot.type_as(keys)
 
 
+# ── Attention ────────────────────────────────────────────────────────────────
+
 class Attention(nn.Module):
     def __init__(self, config):
         super().__init__()
@@ -57,39 +83,58 @@ class Attention(nn.Module):
         self.n_kv_heads  = config.attention_n_kv_heads
         self.batch_size  = config.batch_size
         self.max_seq_len = config.max_seq_len
-        d = config.d_model
+        d             = config.d_model
         self.head_dim = d // self.n_heads
         self.n_rep    = self.n_heads // self.n_kv_heads
-        self.q_proj = nn.Linear(d, self.n_heads    * self.head_dim, bias=False)
-        self.k_proj = nn.Linear(d, self.n_kv_heads * self.head_dim, bias=False)
-        self.v_proj = nn.Linear(d, self.n_kv_heads * self.head_dim, bias=False)
-        self.o_proj = nn.Linear(self.n_heads * self.head_dim, d,    bias=False)
-        self.rope   = RoPE(config)
+        self.q_proj   = nn.Linear(d, self.n_heads    * self.head_dim, bias=False)
+        self.k_proj   = nn.Linear(d, self.n_kv_heads * self.head_dim, bias=False)
+        self.v_proj   = nn.Linear(d, self.n_kv_heads * self.head_dim, bias=False)
+        self.o_proj   = nn.Linear(self.n_heads * self.head_dim, d,    bias=False)
+        self.rope     = RoPE(config)
+
     def forward(self, input, mask=None, past_key_values=None, use_cache=False):
         bsz, seq_len, _ = input.shape
-        q = self.q_proj(input).view(bsz, seq_len, self.n_heads,    self.head_dim)
-        k = self.k_proj(input).view(bsz, seq_len, self.n_kv_heads, self.head_dim)
-        v = self.v_proj(input).view(bsz, seq_len, self.n_kv_heads, self.head_dim)
-        sp = past_key_values[0].shape[1] if past_key_values is not None else 0
-        q, k = self.rope(q, k, sp)
+        queries = self.q_proj(input).view(bsz, seq_len, self.n_heads,    self.head_dim)
+        keys    = self.k_proj(input).view(bsz, seq_len, self.n_kv_heads, self.head_dim)
+        values  = self.v_proj(input).view(bsz, seq_len, self.n_kv_heads, self.head_dim)
+
+        start_pos = past_key_values[0].shape[1] if past_key_values is not None else 0
+        queries, keys = self.rope(queries, keys, start_pos)
+
         if past_key_values is not None:
-            k = torch.cat([past_key_values[0], k], dim=1)
-            v = torch.cat([past_key_values[1], v], dim=1)
-        ck, cv = (k, v) if use_cache else (None, None)
-        q, k, v = q.transpose(1,2), k.transpose(1,2), v.transpose(1,2)
-        gqa = self.n_rep > 1
-        if gqa and q.device.type == "mps":
-            k = k.repeat_interleave(self.n_rep, dim=-3)
-            v = v.repeat_interleave(self.n_rep, dim=-3)
-            gqa = False
+            keys   = torch.cat([past_key_values[0], keys],   dim=1)
+            values = torch.cat([past_key_values[1], values], dim=1)
+
+        cached_keys   = keys   if use_cache else None
+        cached_values = values if use_cache else None
+
+        queries = queries.transpose(1, 2)
+        keys    = keys.transpose(1, 2)
+        values  = values.transpose(1, 2)
+
+        apply_gqa = self.n_rep > 1
+        if apply_gqa and queries.device.type == "mps":
+            keys   = keys.repeat_interleave(self.n_rep, dim=-3)
+            values = values.repeat_interleave(self.n_rep, dim=-3)
+            apply_gqa = False
+
+        # FIX: guard mask against None (happens during generation when seq_len==1)
+        attn_mask = mask.to(queries.dtype) if mask is not None else None
+
         with sdpa_kernel(backends=[SDPBackend.CUDNN_ATTENTION, SDPBackend.MATH]):
-            out = F.scaled_dot_product_attention(
-                q.contiguous(), k.contiguous(), v.contiguous(),
-                attn_mask=mask.to(q.dtype) if mask is not None else None,
-                enable_gqa=gqa,
+            attn_output = F.scaled_dot_product_attention(
+                queries.contiguous(),
+                keys.contiguous(),
+                values.contiguous(),
+                attn_mask=attn_mask,
+                enable_gqa=apply_gqa,
             )
-        return self.o_proj(out.transpose(1,2).contiguous().view(bsz, seq_len, -1)), (ck, cv)
 
+        attn_output = attn_output.transpose(1, 2).contiguous().view(bsz, seq_len, -1)
+        return self.o_proj(attn_output), (cached_keys, cached_values)
+
+
+# ── SwiGLU ───────────────────────────────────────────────────────────────────
 
 class SwiGLU(nn.Module):
     def __init__(self, config):
@@ -97,10 +142,13 @@ class SwiGLU(nn.Module):
         self.w_0 = nn.Linear(config.d_model, config.activation_hidden_dim, bias=False)
         self.w_1 = nn.Linear(config.d_model, config.activation_hidden_dim, bias=False)
         self.w_2 = nn.Linear(config.activation_hidden_dim, config.d_model, bias=False)
+
     def forward(self, x):
         return self.w_2(F.silu(self.w_0(x)) * self.w_1(x))
 
 
+# ── PicoDecoderBlock ─────────────────────────────────────────────────────────
+
 class PicoDecoderBlock(nn.Module):
     def __init__(self, config):
         super().__init__()
@@ -108,58 +156,95 @@ class PicoDecoderBlock(nn.Module):
         self.swiglu         = SwiGLU(config)
         self.attention_norm = RMSNorm(config)
         self.swiglu_norm    = RMSNorm(config)
+
     def forward(self, input, mask=None, past_key_values=None, use_cache=False):
-        a, c = self.attention(self.attention_norm(input), mask=mask,
-                              past_key_values=past_key_values, use_cache=use_cache)
-        h = input + a
-        return h + self.swiglu(self.swiglu_norm(h)), c
+        attention_output, cached_key_values = self.attention(
+            self.attention_norm(input),
+            mask=mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+        )
+        h   = input + attention_output
+        out = h + self.swiglu(self.swiglu_norm(h))
+        return out, cached_key_values
 
 
+# ── PicoDecoder (standalone, used during training) ───────────────────────────
+
 class PicoDecoder(nn.Module):
     def __init__(self, model_config):
         super().__init__()
-        self.config = model_config
+        self.config            = model_config
         self.embedding_proj    = nn.Embedding(model_config.vocab_size, model_config.d_model)
-        self.layers            = nn.ModuleList([PicoDecoderBlock(model_config) for _ in range(model_config.n_layers)])
+        self.layers            = nn.ModuleList(
+            [PicoDecoderBlock(model_config) for _ in range(model_config.n_layers)])
         self.output_norm       = RMSNorm(model_config)
-        self.de_embedding_proj = nn.Linear(model_config.d_model, model_config.vocab_size, bias=False)
+        self.de_embedding_proj = nn.Linear(
+            model_config.d_model, model_config.vocab_size, bias=False)
+
     def convert_to_hf_model(self):
-        hf = PicoDecoderHF(PicoDecoderHFConfig.from_dataclass(self.config))
-        hf.load_state_dict(self.state_dict(prefix="pico_decoder."))
-        return hf
+        hf_config = PicoDecoderHFConfig.from_dataclass(self.config)
+        hf_model  = PicoDecoderHF(hf_config)
+        hf_model.load_state_dict(self.state_dict())
+        return hf_model
+
     def forward(self, input_ids, past_key_values=None, use_cache=False):
-        sl = input_ids.shape[-1]
-        h  = self.embedding_proj(input_ids)
-        sp = 0 if past_key_values is None else past_key_values[0][0].shape[1]
+        seq_len   = input_ids.shape[-1]
+        h         = self.embedding_proj(input_ids)
+        start_pos = 0 if past_key_values is None else past_key_values[0][0].shape[1]
+
         mask = None
-        if sl > 1:
-            mask = torch.triu(torch.full((sl, sl), float("-inf")), diagonal=1)
+        if seq_len > 1:
+            mask = torch.full((seq_len, seq_len), float("-inf"))
+            mask = torch.triu(mask, diagonal=1)
             if past_key_values is not None:
-                mask = torch.hstack([torch.zeros((sl, sp)), mask])
+                mask = torch.hstack([torch.zeros((seq_len, start_pos)), mask])
             mask = mask.to(h.device)
-        ckv = () if use_cache else None
-        for i, layer in enumerate(self.layers):
-            lp = past_key_values[i] if past_key_values is not None else None
-            h, lc = layer(h, mask=mask, past_key_values=lp, use_cache=use_cache)
+
+        cached_key_values = () if use_cache else None
+        for idx, layer in enumerate(self.layers):
+            layer_past = past_key_values[idx] if past_key_values is not None else None
+            h, layer_cached = layer(
+                h, mask=mask, past_key_values=layer_past, use_cache=use_cache)
             if use_cache:
-                ckv += (lc,)
-        return self.de_embedding_proj(self.output_norm(h)).float(), ckv
+                cached_key_values += (layer_cached,)
+
+        h      = self.output_norm(h)
+        logits = self.de_embedding_proj(h).float()
+        return logits, cached_key_values
 
 
-# ── HuggingFace wrappers ──────────────────────────────────────────────────────
+# ── PicoDecoderHFConfig ───────────────────────────────────────────────────────
 
 class PicoDecoderHFConfig(PretrainedConfig):
+    """
+    HuggingFace config for BeetleLM PicoDecoder.
+
+    Defaults match generate_configs.py MODEL_BASE. vocab_size is overridden
+    per-repo in config.json because the trainer sets it from the tokenizer.
+    """
     model_type = "pico_decoder"
 
-    # FIX 1 + 2: explicit __init__ with MODEL_BASE defaults; guards None/0 kv_heads
-    def __init__(self,
-                 n_layers=14, d_model=768, vocab_size=32768,
-                 attention_n_heads=12, attention_n_kv_heads=1,
-                 max_seq_len=512, batch_size=64, position_emb_theta=10000.0,
-                 activation_hidden_dim=3072, norm_eps=1e-5, dropout=0.1,
-                 **kwargs):
-        if not attention_n_kv_heads:          # catches None, 0, missing
+    def __init__(
+        self,
+        # Architecture — defaults from generate_configs.py MODEL_BASE
+        n_layers               = 14,
+        d_model                = 768,
+        vocab_size             = 32768,   # overridden per-repo in config.json
+        attention_n_heads      = 12,
+        attention_n_kv_heads   = 1,       # MQA
+        max_seq_len            = 512,
+        batch_size             = 64,
+        position_emb_theta     = 10000.0,
+        activation_hidden_dim  = 3072,
+        norm_eps               = 1e-5,
+        dropout                = 0.1,
+        **kwargs,
+    ):
+        # FIX: guard against None/0/missing attention_n_kv_heads in old config.json
+        if not attention_n_kv_heads:
             attention_n_kv_heads = attention_n_heads
+
         super().__init__(**kwargs)
         self.n_layers              = n_layers
         self.d_model               = d_model
@@ -177,7 +262,9 @@ class PicoDecoderHFConfig(PretrainedConfig):
     def from_dict(cls, config_dict: Dict[str, Any], **kwargs) -> "PicoDecoderHFConfig":
         pico_config = cls(**config_dict)
         return_unused_kwargs = kwargs.pop("return_unused_kwargs", False)
-        unused_kwargs = {k: v for k, v in kwargs.items() if not hasattr(pico_config, k)}
+        unused_kwargs = {
+            k: v for k, v in kwargs.items() if not hasattr(pico_config, k)
+        }
         if return_unused_kwargs:
             return pico_config, unused_kwargs
         return pico_config
@@ -187,44 +274,96 @@ class PicoDecoderHFConfig(PretrainedConfig):
         return cls.from_dict(asdict(model_config))
 
 
+# ── PicoDecoderHF ─────────────────────────────────────────────────────────────
+
 class PicoDecoderHF(PreTrainedModel):
-    """Load with: AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True)"""
+    """
+    HuggingFace wrapper for BeetleLM PicoDecoder.
+
+    IMPORTANT — weights live at the TOP LEVEL (not under self.pico_decoder)
+    because the trainer saves raw PicoDecoder state dicts:
+      checkpoint keys: embedding_proj.weight, layers.0.attention.q_proj.weight ...
+    A self.pico_decoder wrapper would expect pico_decoder.embedding_proj.weight
+    which does not exist in any saved checkpoint.
+    """
     config_class       = PicoDecoderHFConfig
     _no_split_modules  = ["PicoDecoderBlock", "Attention", "SwiGLU", "RMSNorm"]
-    _tied_weights_keys = []   # FIX 3
+    _tied_weights_keys = []
 
-    # FIX 4: explicit property — transformers >= 4.38 calls this directly
+    # FIX: transformers >= 4.38 calls .keys() on this — must return a dict
     @property
     def all_tied_weights_keys(self):
-        return self._tied_weights_keys
+        return {}
 
     def __init__(self, config: PicoDecoderHFConfig):
         super().__init__(config)
-        self.pico_decoder = PicoDecoder(config)
+        # FIX: top-level storage — matches raw PicoDecoder checkpoint keys
+        self.embedding_proj    = nn.Embedding(config.vocab_size, config.d_model)
+        self.layers            = nn.ModuleList(
+            [PicoDecoderBlock(config) for _ in range(config.n_layers)])
+        self.output_norm       = RMSNorm(config)
+        self.de_embedding_proj = nn.Linear(
+            config.d_model, config.vocab_size, bias=False)
 
     def get_input_embeddings(self):
-        return self.pico_decoder.embedding_proj
+        return self.embedding_proj
 
     def set_input_embeddings(self, value):
-        self.pico_decoder.embedding_proj = value
+        self.embedding_proj = value
+
+    def forward(
+        self,
+        input_ids     = None,
+        past_key_values = None,
+        use_cache     = False,
+        labels        = None,
+        **kwargs,
+    ):
+        seq_len   = input_ids.shape[-1]
+        h         = self.embedding_proj(input_ids)
+        start_pos = 0 if past_key_values is None else past_key_values[0][0].shape[1]
+
+        mask = None
+        if seq_len > 1:
+            mask = torch.full((seq_len, seq_len), float("-inf"))
+            mask = torch.triu(mask, diagonal=1)
+            if past_key_values is not None:
+                mask = torch.hstack([torch.zeros((seq_len, start_pos)), mask])
+            mask = mask.to(h.device)
+
+        cached_key_values = () if use_cache else None
+        for idx, layer in enumerate(self.layers):
+            layer_past = past_key_values[idx] if past_key_values is not None else None
+            h, layer_cached = layer(
+                h, mask=mask, past_key_values=layer_past, use_cache=use_cache)
+            if use_cache:
+                cached_key_values += (layer_cached,)
+
+        logits = self.de_embedding_proj(self.output_norm(h)).float()
 
-    def forward(self, input_ids=None, past_key_values=None,
-                use_cache=False, labels=None, **kwargs):
-        logits, new_past = self.pico_decoder(input_ids, past_key_values, use_cache)
         loss = None
         if labels is not None:
+            shift_logits = logits[:, :-1].contiguous()
+            shift_labels = labels[:, 1:].contiguous().clamp(0, self.config.vocab_size - 1)
             loss = F.cross_entropy(
-                logits[:, :-1].contiguous().view(-1, self.config.vocab_size),
-                labels[:, 1:].contiguous().clamp(0, self.config.vocab_size - 1).view(-1),
+                shift_logits.view(-1, self.config.vocab_size),
+                shift_labels.view(-1),
             )
+
         if use_cache:
-            return CausalLMOutputWithPast(loss=loss, logits=logits, past_key_values=new_past)
+            return CausalLMOutputWithPast(
+                loss=loss, logits=logits, past_key_values=cached_key_values)
         return CausalLMOutput(loss=loss, logits=logits)
 
     def prepare_inputs_for_generation(self, input_ids, past_key_values=None, **kwargs):
-        return {"input_ids": input_ids, "past_key_values": past_key_values, "use_cache": True}
+        return {
+            "input_ids":      input_ids,
+            "past_key_values": past_key_values,
+            "use_cache":      True,
+        }
 
 
+# Auto-class registration
 PicoDecoderHFConfig.register_for_auto_class()
 PicoDecoderHF.register_for_auto_class("AutoModel")
 PicoDecoderHF.register_for_auto_class("AutoModelForCausalLM")