fix: copia esatta architettura da train.py, RoPE senza cast dtype

modeling_quark.py

@@ -1,5 +1,5 @@
 """
-Quark language model — compatibile con AutoModelForCausalLM HuggingFace.
+Quark language model — copia esatta dell'architettura di training.
 """
 import math
 import torch
@@ -7,7 +7,6 @@ import torch.nn as nn
 import torch.nn.functional as F
 from transformers import PreTrainedModel
 from transformers.modeling_outputs import CausalLMOutputWithPast
-
 from .configuration_quark import QuarkConfig
 
 
@@ -25,6 +24,7 @@ class RMSNorm(nn.Module):
 class RotaryEmbedding(nn.Module):
     def __init__(self, head_dim, max_seq_len, theta=10_000.0):
         super().__init__()
+        assert head_dim % 2 == 0
         inv_freq = 1.0 / (theta ** (torch.arange(0, head_dim, 2).float() / head_dim))
         self.register_buffer("inv_freq", inv_freq, persistent=False)
         self._build_cache(max_seq_len)
@@ -46,19 +46,21 @@ class RotaryEmbedding(nn.Module):
         T = q.size(2)
         if T > self._max:
             self._build_cache(T)
-        cos = self.cos_cache[:, :, :T, :].to(dtype=q.dtype)
-        sin = self.sin_cache[:, :, :T, :].to(dtype=q.dtype)
-        return q * cos + self._rotate_half(q) * sin,                k * cos + self._rotate_half(k) * sin
+        cos = self.cos_cache[:, :, :T, :]
+        sin = self.sin_cache[:, :, :T, :]
+        q = q * cos + self._rotate_half(q) * sin
+        k = k * cos + self._rotate_half(k) * sin
+        return q, k
 
 
 class GroupedQueryAttention(nn.Module):
     def __init__(self, cfg):
         super().__init__()
+        assert cfg.n_heads % cfg.n_kv_heads == 0
         self.n_heads    = cfg.n_heads
         self.n_kv_heads = cfg.n_kv_heads
         self.n_groups   = cfg.n_heads // cfg.n_kv_heads
         self.head_dim   = cfg.head_dim
-
         self.q_proj = nn.Linear(cfg.d_model, cfg.n_heads    * cfg.head_dim, bias=cfg.qkv_bias)
         self.k_proj = nn.Linear(cfg.d_model, cfg.n_kv_heads * cfg.head_dim, bias=cfg.qkv_bias)
         self.v_proj = nn.Linear(cfg.d_model, cfg.n_kv_heads * cfg.head_dim, bias=cfg.qkv_bias)
@@ -75,10 +77,6 @@ class GroupedQueryAttention(nn.Module):
         if self.n_groups > 1:
             k = k.repeat_interleave(self.n_groups, dim=1)
             v = v.repeat_interleave(self.n_groups, dim=1)
-        # Forza dtype uniforme prima di SDPA
-        dtype = q.dtype
-        k = k.to(dtype)
-        v = v.to(dtype)
         out = F.scaled_dot_product_attention(
             q, k, v, attn_mask=None,
             dropout_p=self.drop if self.training else 0.0,
@@ -108,7 +106,7 @@ class TransformerBlock(nn.Module):
         self.ffn       = SwiGLUFFN(cfg)
 
     def forward(self, x, **kwargs):
-        x = x + self.attn(self.norm_attn(x), **kwargs)
+        x = x + self.attn(self.norm_attn(x))
         x = x + self.ffn(self.norm_ffn(x))
         return x
 
@@ -116,6 +114,7 @@ class TransformerBlock(nn.Module):
 class QuarkPreTrainedModel(PreTrainedModel):
     config_class      = QuarkConfig
     base_model_prefix = "model"
+    _keys_to_ignore_on_load_missing = ["lm_head.weight"]
     supports_gradient_checkpointing = False
 
     def _init_weights(self, module):
@@ -129,62 +128,38 @@ class QuarkPreTrainedModel(PreTrainedModel):
 
 
 class QuarkForCausalLM(QuarkPreTrainedModel):
-    # lm_head è tied con embed_tokens — non è missing, è intenzionale
     _keys_to_ignore_on_load_missing = ["lm_head.weight"]
-    """
-    Quark autoregressive language model.
-    Compatibile con AutoModelForCausalLM.from_pretrained(..., trust_remote_code=True)
-    """
-    def __init__(self, config: QuarkConfig):
+
+    def __init__(self, config):
         super().__init__(config)
         self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model)
         self.layers       = nn.ModuleList([TransformerBlock(config) for _ in range(config.n_layers)])
         self.norm         = RMSNorm(config.d_model, config.rms_eps)
         self.lm_head      = nn.Linear(config.d_model, config.vocab_size, bias=False)
-
-        if config.tie_word_embeddings:
-            self.lm_head.weight = self.embed_tokens.weight
-
+        self.lm_head.weight = self.embed_tokens.weight
         self.post_init()
 
-    def get_input_embeddings(self):
-        return self.embed_tokens
-
-    def set_input_embeddings(self, value):
-        self.embed_tokens = value
-
-    def get_output_embeddings(self):
-        return self.lm_head
-
-    def set_output_embeddings(self, value):
-        self.lm_head = value
+    def get_input_embeddings(self):  return self.embed_tokens
+    def set_input_embeddings(self, v): self.embed_tokens = v
+    def get_output_embeddings(self): return self.lm_head
+    def set_output_embeddings(self, v): self.lm_head = v
 
     def tie_weights(self, **kwargs):
-        """HF chiama questo metodo dopo il caricamento dei pesi."""
-        if self.config.tie_word_embeddings:
-            self.lm_head.weight = self.embed_tokens.weight
-
-    def forward(
-        self,
-        input_ids      = None,
-        attention_mask = None,
-        labels         = None,
-        **kwargs,
-    ):
+        self.lm_head.weight = self.embed_tokens.weight
+
+    def forward(self, input_ids=None, attention_mask=None, labels=None, **kwargs):
         x = self.embed_tokens(input_ids)
         for layer in self.layers:
-            x = layer(x, attention_mask=attention_mask)
+            x = layer(x)
         x      = self.norm(x)
         logits = self.lm_head(x)
-
-        loss = None
+        loss   = None
         if labels is not None:
             loss = F.cross_entropy(
-                logits[:, :-1, :].contiguous().view(-1, self.config.vocab_size),
+                logits[:, :-1].contiguous().view(-1, self.config.vocab_size),
                 labels[:, 1:].contiguous().view(-1),
                 ignore_index=-100,
             )
-
         return CausalLMOutputWithPast(loss=loss, logits=logits)
 
     @torch.no_grad()
@@ -194,27 +169,22 @@ class QuarkForCausalLM(QuarkPreTrainedModel):
         for _ in range(max_new_tokens):
             out    = self(ctx[:, -self.config.max_seq_len:])
             logits = out.logits[0, -1, :].float()
-            # Greedy se temperature=0 o logits degeneri
-            if temperature <= 0 or torch.isnan(logits).any() or torch.isinf(logits).any():
+            if temperature <= 0:
                 token = logits.argmax().view(1, 1)
             else:
-                logits = logits / temperature
-                # Stabilizza prima del softmax
-                logits = logits - logits.max()
-                probs  = F.softmax(logits, dim=-1)
-                probs  = torch.clamp(probs, min=0.0)
-                # Top-p nucleus
+                logits -= logits.max()
+                logits /= temperature
+                probs   = F.softmax(logits, dim=-1)
                 sorted_p, sorted_i = torch.sort(probs, descending=True)
                 cum_p  = torch.cumsum(sorted_p, dim=-1)
-                remove = (cum_p - sorted_p) > top_p
-                sorted_p = sorted_p.clone()
-                sorted_p[remove] = 0.0
-                total = sorted_p.sum()
+                mask   = (cum_p - sorted_p) > top_p
+                sorted_p[mask] = 0.0
+                total  = sorted_p.sum()
                 if total <= 0:
                     token = sorted_i[0].view(1, 1)
                 else:
                     sorted_p /= total
-                    token = sorted_i[torch.multinomial(sorted_p, 1)].unsqueeze(0).unsqueeze(0)
+                    token = sorted_i[torch.multinomial(sorted_p, 1)].view(1, 1)
             ctx = torch.cat([ctx, token], dim=1)
             if eos_token_id is not None and token.item() == eos_token_id:
                 break