llama-stack 0.0.42__py3-none-any.whl → 0.3.4__py3-none-any.whl

llama_stack/models/llama/llama3/model.py

@@ -0,0 +1,304 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+import math
+
+import fairscale.nn.model_parallel.initialize as fs_init
+import torch
+import torch.nn.functional as F
+from fairscale.nn.model_parallel.layers import (
+    ColumnParallelLinear,
+    RowParallelLinear,
+    VocabParallelEmbedding,
+)
+from torch import nn
+
+from .args import ModelArgs
+
+# **NOTE**: This code is not runnable without installing `torch` and `fairscale`
+# dependencies. These dependencies are not part of the default dependencies
+# (requirements.txt) of the `llama-models` package.
+
+
+class RMSNorm(torch.nn.Module):
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.eps = eps
+        self.weight = nn.Parameter(torch.ones(dim))
+
+    def _norm(self, x):
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+    def forward(self, x):
+        output = self._norm(x.float()).type_as(x)
+        return output * self.weight
+
+
+def apply_scaling(freqs: torch.Tensor) -> torch.Tensor:
+    # Values obtained from grid search
+    scale_factor = 8
+    low_freq_factor = 1
+    high_freq_factor = 4
+    old_context_len = 8192  # original llama3 length
+
+    low_freq_wavelen = old_context_len / low_freq_factor
+    high_freq_wavelen = old_context_len / high_freq_factor
+
+    wavelen = 2 * torch.pi / freqs
+    new_freqs = torch.where(wavelen > low_freq_wavelen, freqs / scale_factor, freqs)
+    smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
+    return torch.where(
+        (wavelen >= high_freq_wavelen) & (wavelen <= low_freq_wavelen),
+        (1 - smooth) * new_freqs / scale_factor + smooth * new_freqs,
+        new_freqs,
+    )
+
+
+def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0, use_scaled: bool = False):
+    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
+    t = torch.arange(end, device=freqs.device, dtype=torch.float32)
+    if use_scaled:
+        freqs = apply_scaling(freqs)
+    freqs = torch.outer(t, freqs)
+    freqs_cis = torch.polar(torch.ones_like(freqs), freqs)  # complex64
+    return freqs_cis
+
+
+def reshape_for_broadcast(freqs_cis: torch.Tensor, x: torch.Tensor):
+    ndim = x.ndim
+    assert 0 <= 1 < ndim
+    assert freqs_cis.shape == (x.shape[1], x.shape[-1])
+    shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)]
+    return freqs_cis.view(*shape)
+
+
+def apply_rotary_emb(
+    xq: torch.Tensor,
+    xk: torch.Tensor,
+    freqs_cis: torch.Tensor,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
+    xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
+    freqs_cis = reshape_for_broadcast(freqs_cis, xq_)
+    xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3)
+    xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3)
+    return xq_out.type_as(xq), xk_out.type_as(xk)
+
+
+def repeat_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """torch.repeat_interleave(x, dim=2, repeats=n_rep)"""
+    bs, slen, n_kv_heads, head_dim = x.shape
+    if n_rep == 1:
+        return x
+    return (
+        x[:, :, :, None, :]
+        .expand(bs, slen, n_kv_heads, n_rep, head_dim)
+        .reshape(bs, slen, n_kv_heads * n_rep, head_dim)
+    )
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.n_kv_heads = args.n_heads if args.n_kv_heads is None else args.n_kv_heads
+        world_size = fs_init.get_model_parallel_world_size()
+        self.n_local_heads = args.n_heads // world_size
+        self.n_local_kv_heads = self.n_kv_heads // world_size
+        self.n_rep = self.n_local_heads // self.n_local_kv_heads
+        self.head_dim = args.dim // args.n_heads
+
+        self.wq = ColumnParallelLinear(
+            args.dim,
+            args.n_heads * self.head_dim,
+            bias=False,
+            gather_output=False,
+            init_method=lambda x: x,
+        )
+        self.wk = ColumnParallelLinear(
+            args.dim,
+            self.n_kv_heads * self.head_dim,
+            bias=False,
+            gather_output=False,
+            init_method=lambda x: x,
+        )
+        self.wv = ColumnParallelLinear(
+            args.dim,
+            self.n_kv_heads * self.head_dim,
+            bias=False,
+            gather_output=False,
+            init_method=lambda x: x,
+        )
+        self.wo = RowParallelLinear(
+            args.n_heads * self.head_dim,
+            args.dim,
+            bias=False,
+            input_is_parallel=True,
+            init_method=lambda x: x,
+        )
+
+        self.cache_k = torch.zeros(
+            (
+                args.max_batch_size,
+                args.max_seq_len,
+                self.n_local_kv_heads,
+                self.head_dim,
+            )
+        )
+        self.cache_v = torch.zeros(
+            (
+                args.max_batch_size,
+                args.max_seq_len,
+                self.n_local_kv_heads,
+                self.head_dim,
+            )
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        start_pos: int,
+        freqs_cis: torch.Tensor,
+        mask: torch.Tensor | None,
+    ):
+        bsz, seqlen, _ = x.shape
+        xq, xk, xv = self.wq(x), self.wk(x), self.wv(x)
+
+        xq = xq.view(bsz, seqlen, self.n_local_heads, self.head_dim)
+        xk = xk.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim)
+        xv = xv.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim)
+
+        xq, xk = apply_rotary_emb(xq, xk, freqs_cis=freqs_cis)
+
+        self.cache_k = self.cache_k.to(xq)
+        self.cache_v = self.cache_v.to(xq)
+
+        self.cache_k[:bsz, start_pos : start_pos + seqlen] = xk
+        self.cache_v[:bsz, start_pos : start_pos + seqlen] = xv
+
+        keys = self.cache_k[:bsz, : start_pos + seqlen]
+        values = self.cache_v[:bsz, : start_pos + seqlen]
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        keys = repeat_kv(keys, self.n_rep)  # (bs, cache_len + seqlen, n_local_heads, head_dim)
+        values = repeat_kv(values, self.n_rep)  # (bs, cache_len + seqlen, n_local_heads, head_dim)
+
+        xq = xq.transpose(1, 2)  # (bs, n_local_heads, seqlen, head_dim)
+        keys = keys.transpose(1, 2)  # (bs, n_local_heads, cache_len + seqlen, head_dim)
+        values = values.transpose(1, 2)  # (bs, n_local_heads, cache_len + seqlen, head_dim)
+        scores = torch.matmul(xq, keys.transpose(2, 3)) / math.sqrt(self.head_dim)
+        if mask is not None:
+            scores = scores + mask  # (bs, n_local_heads, seqlen, cache_len + seqlen)
+        scores = F.softmax(scores.float(), dim=-1).type_as(xq)
+        output = torch.matmul(scores, values)  # (bs, n_local_heads, seqlen, head_dim)
+        output = output.transpose(1, 2).contiguous().view(bsz, seqlen, -1)
+        return self.wo(output)
+
+
+class FeedForward(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        hidden_dim: int,
+        multiple_of: int,
+        ffn_dim_multiplier: float | None,
+    ):
+        super().__init__()
+        hidden_dim = int(2 * hidden_dim / 3)
+        # custom dim factor multiplier
+        if ffn_dim_multiplier is not None:
+            hidden_dim = int(ffn_dim_multiplier * hidden_dim)
+        hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
+
+        self.w1 = ColumnParallelLinear(dim, hidden_dim, bias=False, gather_output=False, init_method=lambda x: x)
+        self.w2 = RowParallelLinear(hidden_dim, dim, bias=False, input_is_parallel=True, init_method=lambda x: x)
+        self.w3 = ColumnParallelLinear(dim, hidden_dim, bias=False, gather_output=False, init_method=lambda x: x)
+
+    def forward(self, x):
+        return self.w2(F.silu(self.w1(x)) * self.w3(x))
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, layer_id: int, args: ModelArgs):
+        super().__init__()
+        self.n_heads = args.n_heads
+        self.dim = args.dim
+        self.head_dim = args.dim // args.n_heads
+        self.attention = Attention(args)
+        self.feed_forward = FeedForward(
+            dim=args.dim,
+            hidden_dim=4 * args.dim,
+            multiple_of=args.multiple_of,
+            ffn_dim_multiplier=args.ffn_dim_multiplier,
+        )
+        self.layer_id = layer_id
+        self.attention_norm = RMSNorm(args.dim, eps=args.norm_eps)
+        self.ffn_norm = RMSNorm(args.dim, eps=args.norm_eps)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        start_pos: int,
+        freqs_cis: torch.Tensor,
+        mask: torch.Tensor | None,
+    ):
+        h = x + self.attention(self.attention_norm(x), start_pos, freqs_cis, mask)
+        out = h + self.feed_forward(self.ffn_norm(h))
+        return out
+
+
+class Transformer(nn.Module):
+    def __init__(self, params: ModelArgs):
+        super().__init__()
+        self.params = params
+        self.vocab_size = params.vocab_size
+        self.n_layers = params.n_layers
+
+        self.tok_embeddings = VocabParallelEmbedding(params.vocab_size, params.dim, init_method=lambda x: x)
+
+        self.layers = torch.nn.ModuleList()
+        for layer_id in range(params.n_layers):
+            self.layers.append(TransformerBlock(layer_id, params))
+
+        self.norm = RMSNorm(params.dim, eps=params.norm_eps)
+        self.output = ColumnParallelLinear(params.dim, params.vocab_size, bias=False, init_method=lambda x: x)
+
+        self.freqs_cis = precompute_freqs_cis(
+            params.dim // params.n_heads,
+            params.max_seq_len * 2,
+            params.rope_theta,
+            params.use_scaled_rope,
+        )
+
+    @torch.inference_mode()
+    def forward(self, tokens: torch.Tensor, start_pos: int):
+        _bsz, seqlen = tokens.shape
+        h = self.tok_embeddings(tokens)
+        self.freqs_cis = self.freqs_cis.to(h.device)
+        freqs_cis = self.freqs_cis[start_pos : start_pos + seqlen]
+
+        mask = None
+        if seqlen > 1:
+            mask = torch.full((seqlen, seqlen), float("-inf"), device=tokens.device)
+
+            mask = torch.triu(mask, diagonal=1)
+
+            # https://github.com/pytorch/pytorch/issues/100005
+            # torch.triu is buggy when the device is mps: filled values are
+            # nan instead of 0.
+            if mask.device.type == torch.device("mps").type:
+                mask = torch.nan_to_num(mask, nan=0.0)
+
+            # When performing key-value caching, we compute the attention scores
+            # only for the new sequence. Thus, the matrix of scores is of size
+            # (seqlen, cache_len + seqlen), and the only masked entries are (i, j) for
+            # j > cache_len + i, since row i corresponds to token cache_len + i.
+            mask = torch.hstack([torch.zeros((seqlen, start_pos), device=tokens.device), mask]).type_as(h)
+
+        for layer in self.layers:
+            h = layer(h, start_pos, freqs_cis, mask)
+        h = self.norm(h)
+        output = self.output(h).float()
+        return output

llama_stack/models/llama/llama3/multimodal/__init__.py

@@ -0,0 +1,12 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# top-level folder for each specific model found within the models/ directory at
+# the top-level of this source tree.

llama_stack/models/llama/llama3/multimodal/encoder_utils.py

@@ -0,0 +1,180 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# top-level folder for each specific model found within the models/ directory at
+# the top-level of this source tree.
+
+# Copyright (c) Meta Platforms, Inc. and its affiliates.
+import math
+
+import torch
+import torch.nn.functional as F
+
+from llama_stack.log import get_logger
+
+from .utils import get_negative_inf_value, to_2tuple
+
+logger = get_logger(name=__name__, category="models::llama")
+
+
+def resize_local_position_embedding(orig_pos_embed, grid_size):
+    """
+    Resize position embedding for vision encoder.
+    Original position embedding is [n_tiles * n_tiles + 1, dim]
+    New position embedding will be [grid_size[0] * grid_size[1] + 1, dim]
+    """
+    new_grid_size = to_2tuple(grid_size)
+    orig_grid_size = to_2tuple(int(math.sqrt(len(orig_pos_embed) - 1)))
+
+    new_pos_emb_tok, new_pos_emb_img = (
+        orig_pos_embed[:1],
+        orig_pos_embed[1:],
+    )
+    logger.info(f"resizing position embedding grid-size from {orig_grid_size} to {new_grid_size}")
+
+    new_pos_emb_img = new_pos_emb_img.reshape(1, orig_grid_size[0], orig_grid_size[1], -1).permute(0, 3, 1, 2)
+
+    new_pos_emb_img = F.interpolate(
+        new_pos_emb_img,
+        size=new_grid_size,
+        mode="bilinear",
+        align_corners=True,
+    )
+    new_pos_emb_img = new_pos_emb_img.permute(0, 2, 3, 1).reshape(1, new_grid_size[0] * new_grid_size[1], -1)[0]
+    new_pos_embed = torch.cat([new_pos_emb_tok, new_pos_emb_img], dim=0)
+    return new_pos_embed
+
+
+def initialize_global_position_embedding_from_local(pos_and_cls_embed, grid_size, x_scale, y_scale):
+    """
+    Takes a local position embedding for vision encoder and uses it
+    to initialize the global position embedding.
+    Input: local position embedding of shape [grid_size[0] * grid_size[1] + 1, dim]
+    Returns: global position embedding of shape [x_scale, y_scale, grid_size[0] * grid_size[1] + 1, dim]
+    Here x_scale and y_scale are the number of tiles along x-axis and y-axis respectively.
+    """
+    pos_embed = pos_and_cls_embed[1:]
+    cls_embed = pos_and_cls_embed[0].view(1, 1, 1, -1)
+    grid_size = to_2tuple(grid_size)
+    new_pos_emb_img = pos_embed.reshape(1, grid_size[0], grid_size[1], -1).permute(0, 3, 1, 2)
+    new_grid_size = (x_scale * grid_size[0], y_scale * grid_size[1])
+    new_pos_emb_img = F.interpolate(
+        new_pos_emb_img,
+        size=new_grid_size,
+        mode="bilinear",
+        align_corners=True,
+    )
+    new_pos_emb_img = new_pos_emb_img.permute(0, 2, 3, 1)
+    new_pos_emb_img = new_pos_emb_img.view(x_scale, grid_size[0], y_scale, grid_size[1], -1)
+    new_pos_emb_img = new_pos_emb_img.permute(0, 2, 1, 3, 4).contiguous()
+    new_pos_emb_img = new_pos_emb_img.reshape(x_scale, y_scale, grid_size[0] * grid_size[1], -1)
+    cls_embed = cls_embed.expand(x_scale, y_scale, -1, -1)
+    pos_and_cls_embed = torch.cat([cls_embed, new_pos_emb_img], dim=2)
+    return pos_and_cls_embed
+
+
+def resize_global_position_embedding(pos_and_cls_embed, grid_size, x_scale, y_scale):
+    """
+    Takes a global position embedding for vision encoder and resizes it to new size.
+    Input: global position embedding of shape [x_old, y_old, old_grid_size[0] * old_grid_size[1] + 1, dim]
+    Returns: global position embedding of shape [x_scale, y_scale, grid_size[0] * grid_size[1] + 1, dim]
+    Here x_scale and y_scale are the number of tiles along x-axis and y-axis respectively.
+    """
+    # first remove cls token
+    pos_embed = pos_and_cls_embed[:, :, 1:]
+    cls_embed = pos_and_cls_embed[:, :, 0].unsqueeze(2)
+
+    xs_old, ys_old, ntok, dim = pos_embed.shape
+    old_grid_size = int(math.sqrt(ntok))
+
+    # move to correct form for interpolation
+    pos_embed = pos_embed.view(xs_old, ys_old, old_grid_size, old_grid_size, dim)
+    pos_embed = pos_embed.permute(0, 2, 1, 3, 4).contiguous()
+    pos_embed = pos_embed.view(xs_old * old_grid_size, ys_old * old_grid_size, dim)
+    pos_embed = pos_embed.unsqueeze(0)
+
+    # interpolate
+    new_size = (grid_size[0] * x_scale, grid_size[1] * y_scale)
+    pos_embed = pos_embed.permute(0, 3, 1, 2)
+    pos_embed_resized = F.interpolate(
+        pos_embed,
+        size=new_size,
+        mode="bilinear",
+        align_corners=True,
+    )
+    pos_embed = pos_embed_resized.permute(0, 2, 3, 1)[0]
+
+    # move it back in place
+    pos_embed = pos_embed.view(x_scale, grid_size[0], y_scale, grid_size[1], dim)
+    pos_embed = pos_embed.permute(0, 2, 1, 3, 4).contiguous()
+    pos_embed = pos_embed.view(x_scale, y_scale, grid_size[0] * grid_size[1], dim)
+
+    # interpolate cls token
+    cls_embed = cls_embed.permute(2, 3, 0, 1)
+    cls_embed_resized = F.interpolate(
+        cls_embed,
+        size=(x_scale, y_scale),
+        mode="bilinear",
+        align_corners=True,
+    )
+    cls_embed = cls_embed_resized.permute(2, 3, 0, 1)
+    # add cls token back in
+    pos_and_cls_embed = torch.cat([cls_embed, pos_embed], dim=2)
+
+    return pos_and_cls_embed
+
+
+def build_encoder_attention_mask(
+    x: torch.Tensor,
+    ar: torch.Tensor,
+    ntok: int,
+    num_chunks: int,
+    n_heads: int,
+):
+    """
+    Build vision encoder attention mask that omits padding tokens.
+    """
+    masks = []
+    for arx in ar:
+        mask_i = torch.ones((num_chunks, x.shape[2], 1), dtype=x.dtype)
+        mask_i[: arx[0] * arx[1], :ntok] = 0
+        mask_i = mask_i.view(num_chunks * x.shape[2], -1)
+        mask_i = mask_i @ mask_i.T * get_negative_inf_value(x.dtype)
+        mask_i = mask_i.unsqueeze(0)
+        masks.append(mask_i)
+    masks = torch.stack(masks).to(x.device).expand(-1, n_heads, -1, -1)
+    return masks
+
+
+def expand_num_tokens_to_mult8(x):
+    num_pad_tokens = 8 - (x.shape[-2] % 8)
+    if num_pad_tokens == 0:
+        return x, 0
+    else:
+        return (
+            torch.cat(
+                [
+                    x,
+                    torch.zeros(
+                        (x.shape[0], x.shape[1], num_pad_tokens, x.shape[-1]),
+                        dtype=x.dtype,
+                        device=x.device,
+                    ),
+                ],
+                dim=-2,
+            ),
+            num_pad_tokens,
+        )
+
+
+def contract_num_tokens_from_mult8(x, num_pad_tokens):
+    if num_pad_tokens == 0:
+        return x
+    return x[:, :, :-num_pad_tokens]