Add pico_decoder.py + auto_map config (main)

config.json

@@ -1,6 +1,6 @@
 {
   "architectures": [
-    "PicoDecoderModel"
+    "PicoDecoderHF"
   ],
   "model_type": "pico_decoder",
   "vocab_size": 32000,
@@ -14,5 +14,18 @@
   "rms_norm_eps": 1e-05,
   "tie_word_embeddings": false,
   "torch_dtype": "float32",
-  "transformers_version": "4.40.0"
+  "transformers_version": "4.48.3",
+  "auto_map": {
+    "AutoConfig": "pico_decoder.PicoDecoderHFConfig",
+    "AutoModelForCausalLM": "pico_decoder.PicoDecoderHF"
+  },
+  "d_model": 768,
+  "n_layers": 14,
+  "attention_n_heads": 6,
+  "attention_n_kv_heads": 0,
+  "activation_hidden_dim": 3072,
+  "max_seq_len": 2048,
+  "norm_eps": 1e-06,
+  "position_emb_theta": 10000.0,
+  "batch_size": 1
 }

pico_decoder.py

@@ -0,0 +1,250 @@
+
+"""
+Pico Decoder — BeetleLM
+Adapted from pico-lm/pico-decoder-tiny (Apache 2.0).
+Load with trust_remote_code=True.
+"""
+
+from typing import 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
+
+
+# ── RMSNorm ───────────────────────────────────────────────────────────────────
+
+class RMSNorm(torch.nn.Module):
+    def __init__(self, config):
+        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
+
+
+# ── 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
+        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)
+
+    def get_freqs_cis(self, input_shape, start_pos, end_pos):
+        _f = self._freqs_cis[start_pos:end_pos]
+        ndim = len(input_shape)
+        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),
+        )
+
+
+# ── Attention ─────────────────────────────────────────────────────────────────
+
+class Attention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.n_heads    = config.attention_n_heads
+        self.n_kv_heads = config.attention_n_kv_heads
+        self.n_rep      = self.n_heads // self.n_kv_heads
+        self.max_seq_len = config.max_seq_len
+        self.batch_size  = config.batch_size
+        d = config.d_model
+        self.head_dim = d // self.n_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)
+
+    def forward(self, x, mask=None, past_key_values=None, use_cache=False):
+        bsz, seq_len, _ = x.shape
+        q = self.q_proj(x).view(bsz, seq_len, self.n_heads,    self.head_dim)
+        k = self.k_proj(x).view(bsz, seq_len, self.n_kv_heads, self.head_dim)
+        v = self.v_proj(x).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
+        q, k = self.rope(q, k, 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)
+        apply_gqa = self.n_rep > 1
+        if apply_gqa and q.device.type == "mps":
+            k = k.repeat_interleave(self.n_rep, dim=-3)
+            v = v.repeat_interleave(self.n_rep, dim=-3)
+            apply_gqa = False
+        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=apply_gqa,
+            )
+        out = out.transpose(1,2).contiguous().view(bsz, seq_len, -1)
+        return self.o_proj(out), (ck, cv)
+
+
+# ── SwiGLU ────────────────────────────────────────────────────────────────────
+
+class SwiGLU(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        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__()
+        self.attention      = Attention(config)
+        self.swiglu         = SwiGLU(config)
+        self.attention_norm = RMSNorm(config)
+        self.swiglu_norm    = RMSNorm(config)
+
+    def forward(self, x, mask=None, past_key_values=None, use_cache=False):
+        attn_out, cached = self.attention(
+            self.attention_norm(x), mask=mask,
+            past_key_values=past_key_values, use_cache=use_cache,
+        )
+        h = x + attn_out
+        return h + self.swiglu(self.swiglu_norm(h)), cached
+
+
+# ── PicoDecoder ───────────────────────────────────────────────────────────────
+
+class PicoDecoder(nn.Module):
+    def __init__(self, model_config):
+        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.output_norm       = RMSNorm(model_config)
+        self.de_embedding_proj = nn.Linear(model_config.d_model, model_config.vocab_size, bias=False)
+
+    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]
+        mask = None
+        if seq_len > 1:
+            mask = torch.triu(torch.full((seq_len, seq_len), float("-inf")), 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_kvs = () 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_kvs += (layer_cached,)
+        return self.de_embedding_proj(self.output_norm(h)).float(), cached_kvs
+
+
+# ── HuggingFace Config ────────────────────────────────────────────────────────
+
+class PicoDecoderHFConfig(PretrainedConfig):
+    model_type = "pico_decoder"
+
+    def __init__(
+        self,
+        vocab_size=32000,
+        d_model=256,
+        n_layers=6,
+        attention_n_heads=8,
+        attention_n_kv_heads=4,
+        activation_hidden_dim=1024,
+        max_seq_len=2048,
+        norm_eps=1e-6,
+        position_emb_theta=10000.0,
+        batch_size=1,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.vocab_size             = vocab_size
+        self.d_model                = d_model
+        self.n_layers               = n_layers
+        self.attention_n_heads      = attention_n_heads
+        self.attention_n_kv_heads   = attention_n_kv_heads
+        self.activation_hidden_dim  = activation_hidden_dim
+        self.max_seq_len            = max_seq_len
+        self.norm_eps               = norm_eps
+        self.position_emb_theta     = position_emb_theta
+        self.batch_size             = batch_size
+
+
+# ── HuggingFace Model ─────────────────────────────────────────────────────────
+
+class PicoDecoderHF(PreTrainedModel):
+    """
+    HuggingFace wrapper for BeetleLM PicoDecoder.
+    Usage: AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True)
+    """
+    config_class        = PicoDecoderHFConfig
+    _no_split_modules   = ["PicoDecoderBlock", "Attention", "SwiGLU", "RMSNorm"]
+
+    def __init__(self, config: PicoDecoderHFConfig):
+        super().__init__(config)
+        self.pico_decoder = PicoDecoder(config)
+
+    def get_input_embeddings(self):
+        return self.pico_decoder.embedding_proj
+
+    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):
+        input_ids = input_ids.clamp(0, self.config.vocab_size - 1)
+        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),
+            )
+        if use_cache:
+            return CausalLMOutputWithPast(loss=loss, logits=logits, past_key_values=new_past)
+        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}
+
+
+# Auto-class registration (runs on trust_remote_code import)
+PicoDecoderHFConfig.register_for_auto_class()
+PicoDecoderHF.register_for_auto_class("AutoModel")
+PicoDecoderHF.register_for_auto_class("AutoModelForCausalLM")