Fix pico_decoder.py: __init__ defaults, ZeroDivisionError, all_tied_weights_keys

pico_decoder.py

@@ -1,14 +1,11 @@
-
 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
@@ -21,49 +18,36 @@ class RMSNorm(torch.nn.Module):
         super().__init__()
         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
 
 
 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
+        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))
-        freqs = torch.outer(torch.arange(seq_len), _freqs)
-        return torch.polar(torch.ones_like(freqs), freqs)
-
+        return torch.polar(torch.ones_like(f := torch.outer(torch.arange(seq_len), _freqs)), f)
     def get_freqs_cis(self, input_shape, start_pos, end_pos):
         _f = self._freqs_cis[start_pos:end_pos]
         ndim = len(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)
-
+        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)])
     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),
-        )
+        return (torch.view_as_real(q_ * fc).flatten(3).type_as(queries),
+                torch.view_as_real(k_ * fc).flatten(3).type_as(keys))
 
 
 class Attention(nn.Module):
@@ -81,35 +65,30 @@ class Attention(nn.Module):
         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
-        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)
+        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)
         if past_key_values is not None:
-            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
+            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
         with sdpa_kernel(backends=[SDPBackend.CUDNN_ATTENTION, SDPBackend.MATH]):
             out = F.scaled_dot_product_attention(
-                queries.contiguous(), keys.contiguous(), values.contiguous(),
-                attn_mask=mask.to(queries.dtype) if mask is not None else None,
-                enable_gqa=apply_gqa,
+                q.contiguous(), k.contiguous(), v.contiguous(),
+                attn_mask=mask.to(q.dtype) if mask is not None else None,
+                enable_gqa=gqa,
             )
-        out = out.transpose(1, 2).contiguous().view(bsz, seq_len, -1)
-        return self.o_proj(out), (cached_keys, cached_values)
+        return self.o_proj(out.transpose(1,2).contiguous().view(bsz, seq_len, -1)), (ck, cv)
 
 
 class SwiGLU(nn.Module):
@@ -118,7 +97,6 @@ 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))
 
@@ -130,14 +108,11 @@ 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):
-        attn_out, cached = self.attention(
-            self.attention_norm(input), mask=mask,
-            past_key_values=past_key_values, use_cache=use_cache,
-        )
-        h = input + attn_out
-        return h + self.swiglu(self.swiglu_norm(h)), cached
+        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
 
 
 class PicoDecoder(nn.Module):
@@ -145,60 +120,45 @@ class PicoDecoder(nn.Module):
         super().__init__()
         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)
-
     def convert_to_hf_model(self):
-        hf_config = PicoDecoderHFConfig.from_dataclass(self.config)
-        hf_model  = PicoDecoderHF(hf_config)
-        hf_model.load_state_dict(self.state_dict(prefix="pico_decoder."))
-        return hf_model
-
+        hf = PicoDecoderHF(PicoDecoderHFConfig.from_dataclass(self.config))
+        hf.load_state_dict(self.state_dict(prefix="pico_decoder."))
+        return hf
     def forward(self, input_ids, past_key_values=None, use_cache=False):
-        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]
+        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]
         mask = None
-        if seq_len > 1:
-            mask = torch.triu(torch.full((seq_len, seq_len), float("-inf")), diagonal=1)
+        if sl > 1:
+            mask = torch.triu(torch.full((sl, sl), float("-inf")), diagonal=1)
             if past_key_values is not None:
-                mask = torch.hstack([torch.zeros((seq_len, start_pos)), mask])
+                mask = torch.hstack([torch.zeros((sl, sp)), 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)
+        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)
             if use_cache:
-                cached_key_values += (layer_cached,)
-        return self.de_embedding_proj(self.output_norm(h)).float(), cached_key_values
+                ckv += (lc,)
+        return self.de_embedding_proj(self.output_norm(h)).float(), ckv
 
 
+# ── HuggingFace wrappers ──────────────────────────────────────────────────────
+
 class PicoDecoderHFConfig(PretrainedConfig):
-    """HuggingFace config for BeetleLM PicoDecoder.
-    Defaults match generate_configs.py MODEL_BASE exactly.
-    """
     model_type = "pico_decoder"
 
-    def __init__(
-        self,
-        n_layers: int = 14,
-        d_model: int = 768,
-        vocab_size: int = 32768,
-        attention_n_heads: int = 12,
-        attention_n_kv_heads: int = 1,
-        max_seq_len: int = 512,
-        batch_size: int = 64,
-        position_emb_theta: float = 10000.0,
-        activation_hidden_dim: int = 3072,
-        norm_eps: float = 1e-5,
-        dropout: float = 0.1,
-        **kwargs,
-    ):
-        # Fix: guard against None/0/missing attention_n_kv_heads in old config.json
-        if not attention_n_kv_heads:
+    # 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
             attention_n_kv_heads = attention_n_heads
         super().__init__(**kwargs)
         self.n_layers              = n_layers
@@ -228,12 +188,15 @@ class PicoDecoderHFConfig(PretrainedConfig):
 
 
 class PicoDecoderHF(PreTrainedModel):
-    """HuggingFace wrapper for BeetleLM PicoDecoder.
-    Load with: AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True)
-    """
+    """Load with: AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True)"""
     config_class       = PicoDecoderHFConfig
     _no_split_modules  = ["PicoDecoderBlock", "Attention", "SwiGLU", "RMSNorm"]
-    _tied_weights_keys = []   # Fix: required by transformers >= 4.38
+    _tied_weights_keys = []   # FIX 3
+
+    # FIX 4: explicit property — transformers >= 4.38 calls this directly
+    @property
+    def all_tied_weights_keys(self):
+        return self._tied_weights_keys
 
     def __init__(self, config: PicoDecoderHFConfig):
         super().__init__(config)
@@ -245,16 +208,14 @@ class PicoDecoderHF(PreTrainedModel):
     def set_input_embeddings(self, value):
         self.pico_decoder.embedding_proj = value
 
-    def forward(self, input_ids=None, past_key_values=None, use_cache=False,
-                labels=None, **kwargs):
+    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(
-                shift_logits.view(-1, self.config.vocab_size),
-                shift_labels.view(-1),
+                logits[:, :-1].contiguous().view(-1, self.config.vocab_size),
+                labels[:, 1:].contiguous().clamp(0, self.config.vocab_size - 1).view(-1),
             )
         if use_cache:
             return CausalLMOutputWithPast(loss=loss, logits=logits, past_key_values=new_past)