veylon_model.py

from __future__ import annotations

import numpy as np
import keras
from keras import layers, ops
import jax

from veylon_attention import flash_splash_attention, decode_swa

try:
    from config import (
        CONTEXT,
        vocab_size as Vocab_size,
        D_MODEL,
        numberoflayers,
        numberofheads,
        d_Latent,
        ffn_mult,
        swa_window,
        num_kv_heads,
        use_moe,
        moe_num_experts,
        moe_top_k,
        USE_REMAT,
        MAX_GEN_TOKENS,
    )
except Exception:
    CONTEXT = 2048
    Vocab_size = 32000
    D_MODEL = 512
    numberoflayers = 8
    numberofheads = 8
    d_Latent = 128
    ffn_mult = 3.5
    swa_window = 1024
    num_kv_heads = 2
    use_moe = False
    moe_num_experts = 8
    moe_top_k = 2
    USE_REMAT = True
    MAX_GEN_TOKENS = 256

# NOTE: mixed precision policy is set by train.py / finetune.py before model
# creation.  Do NOT set it here — importing this module must not have side
# effects.


# ─────────────────────────────────────────────────────────────────────────────
# Layers
# ─────────────────────────────────────────────────────────────────────────────

@keras.saving.register_keras_serializable()
class RMSNorm(layers.Layer):
    def __init__(self, epsilon=1e-5, **kwargs):
        super().__init__(**kwargs)
        self.epsilon = epsilon

    def build(self, input_shape):
        self.weight = self.add_weight(shape=(input_shape[-1],), initializer='ones', name='gamma')

    def call(self, x):
        x_fp32 = ops.cast(x, 'float32')
        rms = ops.sqrt(ops.mean(ops.square(x_fp32), axis=-1, keepdims=True) + self.epsilon)
        out = x_fp32 / rms
        return ops.cast(out, x.dtype) * self.weight

    def get_config(self):
        cfg = super().get_config()
        cfg.update({'epsilon': self.epsilon})
        return cfg


@keras.saving.register_keras_serializable()
class RotaryEmbedding(layers.Layer):
    """
    RoPE with `max_seq_len + gen_headroom` so a full-context prompt can still
    decode up to `gen_headroom` tokens without overflowing the cos/sin table.

    `max_seq_len` is preserved in the config so checkpoints remain compatible —
    only the (trainable=False) table shapes grow.
    """

    def __init__(self, max_seq_len, dim, theta=10000.0, gen_headroom=0, **kwargs):
        super().__init__(**kwargs)
        if dim % 2 != 0:
            raise ValueError('RotaryEmbedding dim must be even.')
        self.max_seq_len = max_seq_len
        self.dim = dim
        self.theta = theta
        self.gen_headroom = int(gen_headroom)
        self.table_size = max_seq_len + self.gen_headroom

    def build(self, input_shape):
        half = self.dim // 2
        inv_freq = 1.0 / (self.theta ** (np.arange(0, self.dim, 2).astype(np.float32) / self.dim))
        positions = np.arange(self.table_size, dtype=np.float32)
        freqs = positions[:, None] * inv_freq[None, :]
        self.cos = self.add_weight(
            shape=(self.table_size, half),
            initializer=keras.initializers.Constant(np.cos(freqs)),
            trainable=False, dtype='float32', name='cos_table',
        )
        self.sin = self.add_weight(
            shape=(self.table_size, half),
            initializer=keras.initializers.Constant(np.sin(freqs)),
            trainable=False, dtype='float32', name='sin_table',
        )

    def call(self, x, offset=0):
        seq_len = x.shape[1]
        if seq_len is None:
            seq_len = ops.shape(x)[1]
        half = self.dim // 2
        if offset + seq_len > self.table_size:
            raise ValueError(
                f'RoPE table too small: offset={offset}, seq_len={seq_len}, '
                f'max={self.table_size} (CONTEXT={self.max_seq_len} + gen_headroom={self.gen_headroom})'
            )
        cos = self.cos[offset:offset + seq_len, :half]
        sin = self.sin[offset:offset + seq_len, :half]
        cos = ops.cast(ops.reshape(cos, (1, seq_len, 1, half)), x.dtype)
        sin = ops.cast(ops.reshape(sin, (1, seq_len, 1, half)), x.dtype)
        x1 = x[..., :half]
        x2 = x[..., half:]
        return ops.concatenate([x1 * cos - x2 * sin, x1 * sin + x2 * cos], axis=-1)

    def get_config(self):
        cfg = super().get_config()
        cfg.update({
            'max_seq_len': self.max_seq_len,
            'dim': self.dim,
            'theta': self.theta,
            'gen_headroom': self.gen_headroom,
        })
        return cfg


@keras.saving.register_keras_serializable()
class SwiGLUFFN(layers.Layer):
    def __init__(self, d_model, hidden_mult=3.5, **kwargs):
        super().__init__(**kwargs)
        self.d_model_arg = d_model
        self.hidden_mult = hidden_mult
        self.hidden_dim = int(d_model * hidden_mult * 2 / 3)
        self.hidden_dim = ((self.hidden_dim + 63) // 64) * 64

    def build(self, input_shape):
        d_model = input_shape[-1]
        self.gate_up_proj = self.add_weight(shape=(d_model, 2 * self.hidden_dim), initializer='glorot_uniform', name='gate_up_proj')
        self.down_proj = self.add_weight(shape=(self.hidden_dim, d_model), initializer='glorot_uniform', name='down_proj')

    def call(self, x, training=False):
        gate_up = ops.matmul(x, self.gate_up_proj)
        gate, up = ops.split(gate_up, 2, axis=-1)
        return ops.matmul(ops.silu(gate) * up, self.down_proj)

    def get_config(self):
        cfg = super().get_config()
        cfg.update({'d_model': self.d_model_arg, 'hidden_mult': self.hidden_mult})
        return cfg


@keras.saving.register_keras_serializable()
class MoE_FFN(layers.Layer):
    """
    Mixture-of-Experts FFN (top-k routing, masked dispatch).

    For each routing slot k in [0, top_k) we run each expert once over the
    full batch, zeroing inputs the expert was not selected for, then accumulate
    the gated expert output.  The load-balancing loss uses top-1 argmax routing
    for the `f` term (standard Switch-Transformer formulation); the previous
    code used top-k membership which over-counted and produced the wrong loss.
    """

    def __init__(self, d_model, num_experts=8, top_k=2, hidden_mult=3.5, **kwargs):
        super().__init__(**kwargs)
        self.d_model_arg = d_model
        self.num_experts = num_experts
        self.top_k = top_k
        self.hidden_mult = hidden_mult
        self.experts = [SwiGLUFFN(d_model, hidden_mult) for _ in range(num_experts)]
        self.router = layers.Dense(num_experts, use_bias=False)

    def _load_balancing_loss(self, router_logits, top_idx):
        # Standard Switch-Transformer aux loss.  `f` uses the argmax routing
        # decision (top_idx[..., 0]); `p` uses the router softmax mean.
        router_probs = ops.softmax(router_logits, axis=-1)
        mask = ops.one_hot(top_idx[..., 0], self.num_experts)        # [..., num_experts]
        f = ops.mean(ops.cast(mask, 'float32'), axis=tuple(range(mask.ndim - 1)))
        p = ops.mean(router_probs, axis=tuple(range(router_probs.ndim - 1)))
        return self.num_experts * ops.sum(f * p)

    def call(self, x, training=False):
        # x: [B, S, D]
        B = ops.shape(x)[0]
        S = ops.shape(x)[1]
        D = x.shape[-1]

        router_logits = self.router(x)                              # [B, S, E]
        top_logits, top_idx = ops.top_k(router_logits, self.top_k)  # [B, S, k]
        top_weights = ops.softmax(top_logits, axis=-1)              # [B, S, k]

        if training:
            self.add_loss(self._load_balancing_loss(router_logits, top_idx))

        out = ops.zeros_like(x)
        for kk in range(self.top_k):
            idx_k = top_idx[..., kk]                                 # [B, S]
            w_k = top_weights[..., kk][..., None]                    # [B, S, 1]
            for e in range(self.num_experts):
                mask_e = ops.cast(idx_k == e, x.dtype)[..., None]    # [B, S, 1]
                # mask_e is always a tensor; skip the useless ops.is_tensor guard
                # (it never triggered — removed FIX BUG11)
                # Only run the expert where it is actually selected.
                x_e = x * mask_e
                expert_out = self.experts[e](x_e, training=training)
                out = out + expert_out * w_k * mask_e
        return out

    def get_config(self):
        cfg = super().get_config()
        cfg.update({
            'd_model': self.d_model_arg, 'num_experts': self.num_experts,
            'top_k': self.top_k, 'hidden_mult': self.hidden_mult,
        })
        return cfg


@keras.saving.register_keras_serializable()
class MLAttention(layers.Layer):
    def __init__(self, d_model, n_heads, d_latent, max_seq_len, num_kv_heads=2,
                 swa_window=1024, attn_dropout=0.0, gen_headroom=0, **kwargs):
        super().__init__(**kwargs)
        if d_model % n_heads != 0:
            raise ValueError('d_model must be divisible by n_heads')
        if n_heads % num_kv_heads != 0:
            raise ValueError('n_heads must be divisible by num_kv_heads')
        self.d_model = d_model
        self.n_heads = n_heads
        self.num_kv_heads = num_kv_heads
        self.group_size = n_heads // num_kv_heads
        self.d_head = d_model // n_heads
        self.d_latent = d_latent
        self.max_seq_len = max_seq_len
        self.swa_window = swa_window
        self.gen_headroom = int(gen_headroom)
        self.dropout = layers.Dropout(attn_dropout)
        self.rope = RotaryEmbedding(max_seq_len, d_model // n_heads, gen_headroom=int(gen_headroom))  # FIX BUG10: must be in __init__ for Keras tracking

    def build(self, input_shape):
        self.W_qc = self.add_weight(shape=(self.d_model, self.d_model + self.d_latent), initializer='glorot_uniform', name='W_qc')
        self.W_kv = self.add_weight(shape=(self.d_latent, self.num_kv_heads * 2 * self.d_head), initializer='glorot_uniform', name='W_kv')
        self.W_o = self.add_weight(shape=(self.d_model, self.d_model), initializer='glorot_uniform', name='Wo')

    def _project_kv(self, c):
        kv = ops.matmul(c, self.W_kv)
        return ops.split(kv, 2, axis=-1)

    def call(self, x, training=False):
        B = ops.shape(x)[0]
        S = ops.shape(x)[1]
        qc = ops.matmul(x, self.W_qc)
        q_proj, c = ops.split(qc, [self.d_model], axis=-1)
        q = ops.reshape(q_proj, (B, S, self.n_heads, self.d_head))
        q = self.rope(q, offset=0)
        q = ops.transpose(q, (0, 2, 1, 3))
        k, v = self._project_kv(c)
        k = ops.reshape(k, (B, S, self.num_kv_heads, self.d_head))
        v = ops.reshape(v, (B, S, self.num_kv_heads, self.d_head))
        k = self.rope(k, offset=0)
        k = ops.transpose(k, (0, 2, 1, 3))
        v = ops.transpose(v, (0, 2, 1, 3))

        # Attention remat is REQUIRED during training: it is the per-block
        # jax.checkpoint(process_block) that stops XLA from staging the hidden
        # [S, B, Hkv, W, D] buffer (≈8.6 GB here) inside the fori_loop.  Block-
        # level remat does NOT prevent that forward-time allocation, so this
        # must follow `training` regardless of config.USE_REMAT.
        out = flash_splash_attention(
            q, k, v,
            window_size=min(self.swa_window, self.max_seq_len),
            backend=jax.default_backend(),
            use_gqa=True,
            use_remat=training,
        )
        out = ops.transpose(out, (0, 2, 1, 3))
        out = ops.reshape(out, (B, S, self.d_model))
        out = self.dropout(out, training=training)
        return ops.matmul(out, self.W_o)

    def generate_step(self, x, cache_k=None, cache_v=None, cache_pos=0):
        B = ops.shape(x)[0]
        S = ops.shape(x)[1]

        qc = ops.matmul(x, self.W_qc)
        q_proj, c = ops.split(qc, [self.d_model], axis=-1)

        q = ops.reshape(q_proj, (B, S, self.n_heads, self.d_head))
        q = self.rope(q, offset=cache_pos)
        q = ops.transpose(q, (0, 2, 1, 3))

        k, v = self._project_kv(c)
        k = ops.reshape(k, (B, S, self.num_kv_heads, self.d_head))
        v = ops.reshape(v, (B, S, self.num_kv_heads, self.d_head))
        k = self.rope(k, offset=cache_pos)
        k = ops.transpose(k, (0, 2, 1, 3))
        v = ops.transpose(v, (0, 2, 1, 3))

        if cache_k is None:
            # Prefill: no remat (inference only), no window trim needed.
            out = flash_splash_attention(
                q, k, v,
                window_size=min(self.swa_window, self.max_seq_len),
                backend=jax.default_backend(),
                use_gqa=True,
                use_remat=False,
            )
            new_k = k[:, :, -self.swa_window:, :]
            new_v = v[:, :, -self.swa_window:, :]
        else:
            if S != 1:
                raise ValueError(f'generate_step with cache expects S=1, got S={S}')

            k = ops.concatenate([cache_k, k], axis=2)
            v = ops.concatenate([cache_v, v], axis=2)

            k = k[:, :, -self.swa_window:, :]
            v = v[:, :, -self.swa_window:, :]

            out = decode_swa(q, k, v)
            new_k = k
            new_v = v

        out = ops.transpose(out, (0, 2, 1, 3))
        out = ops.reshape(out, (B, S, self.d_model))
        out = ops.matmul(out, self.W_o)
        return out, new_k, new_v

    def get_config(self):
        cfg = super().get_config()
        cfg.update({
            'd_model': self.d_model, 'n_heads': self.n_heads,
            'num_kv_heads': self.num_kv_heads, 'd_latent': self.d_latent,
            'max_seq_len': self.max_seq_len, 'swa_window': self.swa_window,
            'attn_dropout': self.dropout.rate, 'gen_headroom': self.gen_headroom,
        })
        return cfg


@keras.saving.register_keras_serializable()
class TransformerBlock(layers.Layer):
    def __init__(self, d_model, n_heads, d_latent, ffn_layer, max_seq_len,
                 num_kv_heads=2, swa_window=1024, use_remat=True, gen_headroom=0, **kwargs):
        super().__init__(**kwargs)
        self.d_model = d_model
        self.n_heads = n_heads
        self.d_latent = d_latent
        self.max_seq_len = max_seq_len
        self.num_kv_heads = num_kv_heads
        self.swa_window = swa_window
        self.use_remat = use_remat
        self.gen_headroom = int(gen_headroom)
        self.ffn = keras.saving.deserialize_keras_object(ffn_layer) if isinstance(ffn_layer, dict) else ffn_layer
        self.norm1 = RMSNorm()
        self.norm2 = RMSNorm()
        self.attn = MLAttention(
            d_model, n_heads, d_latent, max_seq_len,
            num_kv_heads=num_kv_heads, swa_window=swa_window, gen_headroom=self.gen_headroom,
        )

    def call(self, x, training=False):
        # Block-level activation/gradient checkpointing (optional, via
        # config.USE_REMAT).  NOTE: this is independent of attention's own
        # remat, which MUST stay on whenever training=True — that's what stops
        # XLA staging the hidden [S,B,Hkv,W,D] buffer (the OOM we hit).
        if training and self.use_remat:
            def _fwd(x_in):
                attn_out = self.attn(self.norm1(x_in), training=True)
                ffn_out = self.ffn(self.norm2(x_in), training=True)
                return attn_out + ffn_out
            return x + jax.checkpoint(_fwd)(x)
        a = self.attn(self.norm1(x), training=training)
        f = self.ffn(self.norm2(x), training=training)
        return x + a + f

    def generate_step(self, x, cache_k=None, cache_v=None, cache_pos=0):
        # Must mirror call()'s PARALLEL residual structure exactly:
        # attn and ffn both read from norm(x) computed on the SAME
        # pre-block x, then both get added to that same original x.
        # Previously this was written as a sequential block (x=x+attn;
        # x=x+ffn(norm2(x))) — ffn ended up conditioned on norm2(x+attn_out)
        # instead of norm2(x), an input distribution the FFN weights were
        # never trained on. That silently produced a different model at
        # generation time than at training time: teacher-forced loss (via
        # call()) never exercises this path and looked fine, while actual
        # autoregressive generation compounded the mismatch every step.
        attn_out, nck, ncv = self.attn.generate_step(
            self.norm1(x), cache_k=cache_k, cache_v=cache_v, cache_pos=cache_pos,
        )
        ffn_out = self.ffn(self.norm2(x), training=False)
        x = x + attn_out + ffn_out
        return x, nck, ncv

    def get_config(self):
        cfg = super().get_config()
        cfg.update({
            'd_model': self.d_model, 'n_heads': self.n_heads, 'd_latent': self.d_latent,
            'ffn_layer': keras.saving.serialize_keras_object(self.ffn),
            'max_seq_len': self.max_seq_len, 'num_kv_heads': self.num_kv_heads,
            'swa_window': self.swa_window, 'use_remat': self.use_remat,
            'gen_headroom': self.gen_headroom,
        })
        return cfg


@keras.saving.register_keras_serializable()
class VeylonModel(keras.Model):
    def __init__(self, vocab_size, d_model, n_layers, n_heads, d_latent, ffn_mult,
                 max_seq_len, use_moe=False, moe_num_experts=8, moe_top_k=2,
                 num_kv_heads=2, swa_window=1024, use_remat=True, gen_headroom=0, **kwargs):
        super().__init__(**kwargs)
        self.vocab_size = vocab_size
        self.d_model = d_model
        self.n_layers = n_layers
        self.n_heads = n_heads
        self.d_latent = d_latent
        self.ffn_mult = ffn_mult
        self.max_seq_len = max_seq_len
        self.use_moe = use_moe
        self.moe_num_experts = moe_num_experts
        self.moe_top_k = moe_top_k
        self.num_kv_heads = num_kv_heads
        self.swa_window = swa_window
        self.use_remat = use_remat
        self.gen_headroom = int(gen_headroom)
        self.embedding = layers.Embedding(vocab_size, d_model, name='token_embedding')
        self.blocks = []
        for i in range(n_layers):
            ffn = (MoE_FFN(d_model, moe_num_experts, moe_top_k, ffn_mult)
                   if use_moe else SwiGLUFFN(d_model, ffn_mult))
            self.blocks.append(TransformerBlock(
                d_model, n_heads, d_latent, ffn, max_seq_len,
                num_kv_heads=num_kv_heads, swa_window=swa_window,
                use_remat=use_remat, gen_headroom=self.gen_headroom, name=f'block_{i}',
            ))
        self.norm = RMSNorm()

    def call(self, inputs, training=False):
        x = self.embedding(inputs)
        for block in self.blocks:
            x = block(x, training=training)
        x = self.norm(x)
        embedding_weights = self.embedding.embeddings
        logits = ops.matmul(x, ops.transpose(embedding_weights))
        return ops.cast(logits, 'float32')

    def generate_step(self, inputs, cache_k=None, cache_v=None, cache_pos=0):
        x = self.embedding(inputs)
        new_cache_k = []
        new_cache_v = []

        if cache_k is None:
            cache_k = [None] * len(self.blocks)
            cache_v = [None] * len(self.blocks)

        for i, block in enumerate(self.blocks):
            x, nck, ncv = block.generate_step(
                x, cache_k=cache_k[i], cache_v=cache_v[i], cache_pos=cache_pos,
            )
            new_cache_k.append(nck)
            new_cache_v.append(ncv)

        x = self.norm(x)
        embedding_weights = self.embedding.embeddings
        logits = ops.matmul(x, ops.transpose(embedding_weights))
        logits = ops.cast(logits, 'float32')
        return logits, new_cache_k, new_cache_v

    def get_config(self):
        cfg = super().get_config()
        cfg.update({
            'vocab_size': self.vocab_size, 'd_model': self.d_model,
            'n_layers': self.n_layers, 'n_heads': self.n_heads,
            'd_latent': self.d_latent, 'ffn_mult': self.ffn_mult,
            'max_seq_len': self.max_seq_len, 'use_moe': self.use_moe,
            'moe_num_experts': self.moe_num_experts, 'moe_top_k': self.moe_top_k,
            'num_kv_heads': self.num_kv_heads, 'swa_window': self.swa_window,
            'use_remat': self.use_remat, 'gen_headroom': self.gen_headroom,
        })
        return cfg


# ─────────────────────────────────────────────────────────────────────────────
# Factory
# ─────────────────────────────────────────────────────────────────────────────

def create_llm(
    vocab_size=Vocab_size, d_model=D_MODEL, n_layers=numberoflayers, n_heads=numberofheads,
    d_latent=d_Latent, ffn_mult=ffn_mult, max_seq_len=CONTEXT, use_moe=use_moe,
    moe_num_experts=moe_num_experts, moe_top_k=moe_top_k, num_kv_heads=num_kv_heads,
    swa_window=swa_window, use_remat=USE_REMAT, gen_headroom=MAX_GEN_TOKENS,
):
    """
    Build a VeylonModel.

    `gen_headroom` sizes the RoPE table to (max_seq_len + gen_headroom) so that
    a full-length prompt can still generate `gen_headroom` tokens during
    inference without "RoPE table too small" errors.  Defaults to MAX_GEN_TOKENS.
    """
    return VeylonModel(
        vocab_size=vocab_size, d_model=d_model, n_layers=n_layers, n_heads=n_heads,
        d_latent=d_latent, ffn_mult=ffn_mult, max_seq_len=max_seq_len, use_moe=use_moe,
        moe_num_experts=moe_num_experts, moe_top_k=moe_top_k, num_kv_heads=num_kv_heads,
        swa_window=swa_window, use_remat=use_remat, gen_headroom=gen_headroom,
    )