ai-edge-torch-nightly 0.2.0.dev20240714__py3-none-any.whl → 0.3.0.dev20240926__py3-none-any.whl

ai_edge_torch/generative/examples/gemma/{gemma.py → gemma1.py}

@@ -12,43 +12,38 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-# Example of building a Gemma model.
 
-import os
-from pathlib import Path
+"""Example of building a Gemma1 model."""
 
-import numpy as np
-import torch
-import torch.nn as nn
-
-from ai_edge_torch.generative.layers.attention import TransformerBlock
+from ai_edge_torch.generative.layers import attention
+from ai_edge_torch.generative.layers import builder
+from ai_edge_torch.generative.layers import kv_cache as kv_utils
 import ai_edge_torch.generative.layers.attention_utils as attn_utils
-import ai_edge_torch.generative.layers.builder as builder
 import ai_edge_torch.generative.layers.model_config as cfg
 import ai_edge_torch.generative.utilities.loader as loading_utils
+import torch
+from torch import nn
 
 TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
     ff_up_proj="model.layers.{}.mlp.up_proj",
     ff_down_proj="model.layers.{}.mlp.down_proj",
     ff_gate_proj="model.layers.{}.mlp.gate_proj",
-    attn_query_proj="model.layers.{}.self_attn.q_proj",
-    attn_key_proj="model.layers.{}.self_attn.k_proj",
-    attn_value_proj="model.layers.{}.self_attn.v_proj",
+    attn_fused_qkv_proj="model.layers.{}.self_attn.qkv_proj",
     attn_output_proj="model.layers.{}.self_attn.o_proj",
     pre_attn_norm="model.layers.{}.input_layernorm",
-    pre_ff_norm="model.layers.{}.post_attention_layernorm",
-    embedding="model.embed_tokens",
+    post_attn_norm="model.layers.{}.post_attention_layernorm",
+    embedding="embedder",
     final_norm="model.norm",
     lm_head=None,
 )
 
 
 class Gemma(nn.Module):
+  """A Gemma model built from the Edge Generative API layers."""
 
   def __init__(self, config: cfg.ModelConfig):
     super().__init__()
 
-    self.config = config
     # Construct model layers.
     self.tok_embedding = nn.Embedding(
         config.vocab_size, config.embedding_dim, padding_idx=0
@@ -60,35 +55,48 @@ class Gemma(nn.Module):
     )
     # Gemma re-uses the embedding as the head projection layer.
     self.lm_head.weight.data = self.tok_embedding.weight.data
+    # Gemma has only one block config.
+    block_config = config.block_config(0)
     self.transformer_blocks = nn.ModuleList(
-        TransformerBlock(config) for _ in range(config.num_layers)
+        attention.TransformerBlock(block_config, config)
+        for _ in range(config.num_layers)
     )
     self.final_norm = builder.build_norm(
         config.embedding_dim,
         config.final_norm_config,
     )
+    attn_config = block_config.attn_config
     self.rope_cache = attn_utils.build_rope_cache(
         size=config.kv_cache_max,
-        dim=int(config.attn_config.rotary_percentage * config.head_dim),
+        dim=int(attn_config.rotary_percentage * attn_config.head_dim),
         base=10_000,
         condense_ratio=1,
         dtype=torch.float32,
         device=torch.device("cpu"),
     )
     self.mask_cache = attn_utils.build_causal_mask_cache(
-        size=config.kv_cache_max, dtype=torch.float32, device=torch.device("cpu")
+        size=config.kv_cache_max,
+        dtype=torch.float32,
+        device=torch.device("cpu"),
     )
     self.config = config
 
-  # The model's forward function takes in additional k/v cache tensors
-  # and returns the updated k/v cache tensors to the caller.
-  # This can be eliminated if we handle k/v cache updates inside the model itself.
   @torch.inference_mode
-  def forward(self, idx: torch.Tensor, input_pos: torch.Tensor) -> torch.Tensor:
-    B, T = idx.size()
-    assert (
-        self.config.max_seq_len >= T
-    ), f"Cannot forward sequence of length {T}, max seq length is only {self.config.max_seq_len}"
+  def forward(
+      self,
+      tokens: torch.Tensor,
+      input_pos: torch.Tensor,
+      kv_cache: kv_utils.KVCache,
+  ) -> dict[torch.Tensor, kv_utils.KVCache]:
+    _, seq_len = tokens.size()
+    assert self.config.max_seq_len >= seq_len, (
+        f"Cannot forward sequence of length {seq_len}, max seq length is only"
+        f" {self.config.max_seq_len}"
+    )
+    assert len(self.transformer_blocks) == len(kv_cache.caches), (
+        "The number of transformer blocks and the number of KV cache entries"
+        " must be the same."
+    )
 
     cos, sin = self.rope_cache
     cos = cos.index_select(0, input_pos)
@@ -97,20 +105,35 @@ class Gemma(nn.Module):
     mask = mask[:, :, :, : self.config.kv_cache_max]
 
     # token embeddings of shape (b, t, n_embd)
-    x = self.tok_embedding(idx)
+    x = self.tok_embedding(tokens)
     x = x * (self.config.embedding_dim**0.5)
 
+    updated_kv_entires = []
     for i, block in enumerate(self.transformer_blocks):
-      x = block(x, (cos, sin), mask, input_pos)
+      kv_entry = kv_cache.caches[i] if kv_cache else None
+      x, kv_entry = block(x, (cos, sin), mask, input_pos, kv_entry)
+      if kv_entry:
+        updated_kv_entires.append(kv_entry)
+    updated_kv_cache = kv_utils.KVCache(tuple(updated_kv_entires))
 
     x = self.final_norm(x)
-    res = self.lm_head(x)  # (b, t, vocab_size)
-    return res
+    logits = self.lm_head(x)  # (b, t, vocab_size)
+    return {"logits": logits, "kv_cache": updated_kv_cache}
 
 
 def get_model_config_2b(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for a Gemma 2B model.
+
+  Args:
+    kv_cache_max_len (int): The maximum sequence length of the KV cache. Default
+      is 1024.
+
+  Returns:
+    The model config for a Gemma 2B model.
+  """
   attn_config = cfg.AttentionConfig(
       num_heads=8,
+      head_dim=256,
       num_query_groups=1,
       rotary_percentage=1.0,
   )
@@ -124,51 +147,42 @@ def get_model_config_2b(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       epsilon=1e-6,
       zero_centered=True,
   )
+  block_config = cfg.TransformerBlockConfig(
+      attn_config=attn_config,
+      ff_config=ff_config,
+      pre_attention_norm_config=norm_config,
+      post_attention_norm_config=norm_config,
+  )
   config = cfg.ModelConfig(
       vocab_size=256000,
       num_layers=18,
       max_seq_len=8192,
       embedding_dim=2048,
       kv_cache_max_len=kv_cache_max_len,
-      attn_config=attn_config,
-      ff_config=ff_config,
-      pre_attention_norm_config=norm_config,
-      pre_ff_norm_config=norm_config,
+      block_configs=block_config,
       final_norm_config=norm_config,
-      parallel_residual=False,
       lm_head_use_bias=False,
       enable_hlfb=True,
   )
   return config
 
 
-def get_fake_model_config_2b_for_test() -> cfg.ModelConfig:
-  config = get_model_config_2b()
+def get_fake_model_config(kv_cache_max_len: int = 128) -> cfg.ModelConfig:
+  config = get_model_config_2b(kv_cache_max_len)
+  # Gemma has only one block config.
+  config.block_config(0).ff_config.intermediate_size = 128
+  config.vocab_size = 128
   config.num_layers = 2
+  config.max_seq_len = 2 * kv_cache_max_len
   return config
 
 
-def build_2b_model(checkpoint_path, **kwargs) -> nn.Module:
+def build_2b_model(checkpoint_path: str, **kwargs) -> nn.Module:
   config = get_model_config_2b(**kwargs)
   model = Gemma(config)
   loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
-  # since embedding and lm-head use the same weight, we need to set strict
+  # Since embedding and lm-head use the same weight, we need to set strict
   # to False.
   loader.load(model, strict=False)
+  model.eval()
   return model
-
-
-def define_and_run_2b() -> None:
-  kv_cache_max_len = 1024
-  checkpoint_path = os.path.join(Path.home(), "Downloads/llm_data/gemma-2b")
-  model = build_2b_model(checkpoint_path, kv_cache_max_len=kv_cache_max_len)
-  idx = torch.from_numpy(np.array([[1, 2, 3, 4]]))
-  tokens = torch.full((1, kv_cache_max_len), 0, dtype=torch.long, device="cpu")
-  tokens[0, :4] = idx
-  input_pos = torch.arange(0, kv_cache_max_len)
-  print("running an inference")
-  print(model.forward(tokens, input_pos))
-
-
-if __name__ == "__main__":
-  define_and_run_2b()

ai_edge_torch/generative/examples/gemma/gemma2.py

@@ -0,0 +1,267 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+"""Example of building a Gemma2 model."""
+
+import os
+from typing import Optional, Tuple
+
+from ai_edge_torch.generative.layers import attention
+from ai_edge_torch.generative.layers import builder
+from ai_edge_torch.generative.layers import kv_cache as kv_utils
+import ai_edge_torch.generative.layers.attention_utils as attn_utils
+import ai_edge_torch.generative.layers.model_config as cfg
+import ai_edge_torch.generative.utilities.loader as loading_utils
+import torch
+from torch import nn
+
+TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
+    ff_up_proj="model.layers.{}.mlp.up_proj",
+    ff_down_proj="model.layers.{}.mlp.down_proj",
+    ff_gate_proj="model.layers.{}.mlp.gate_proj",
+    attn_fused_qkv_proj="model.layers.{}.self_attn.qkv_proj",
+    attn_output_proj="model.layers.{}.self_attn.o_proj",
+    pre_attn_norm="model.layers.{}.input_layernorm",
+    post_attn_norm="model.layers.{}.post_attention_layernorm",
+    pre_ff_norm="model.layers.{}.pre_feedforward_layernorm",
+    post_ff_norm="model.layers.{}.post_feedforward_layernorm",
+    embedding="embedder",
+    final_norm="model.norm",
+    lm_head=None,
+)
+
+
+class Gemma2Block(attention.TransformerBlock):
+
+  def forward(
+      self,
+      x: torch.Tensor,
+      rope: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+      mask: Optional[torch.Tensor] = None,
+      input_pos: Optional[torch.Tensor] = None,
+      kv_cache: kv_utils.KVCacheEntry = None,
+  ) -> Tuple[torch.Tensor, Optional[kv_utils.KVCacheEntry]]:
+    """Forward function of the Gemma2Block.
+
+    Exactly the same as TransformerBlock but we call the post-attention norm
+    immediately after attention and not after the residual pointwise addition.
+
+    Args:
+      x (torch.Tensor): the input tensor.
+      rope (Tuple[torch.Tensor, torch.Tensor]): the input rope tensor.
+      mask (torch.Tensor): the optional mask tensor.
+      input_pos (torch.Tensor): the optional input position tensor.
+      kv_cache (KVCacheEntry): the optional kv cache entry.
+
+    Returns:
+      output activation from this transformer block, and updated kv cache (if
+      passed in).
+    """
+
+    x_norm = self.pre_atten_norm(x)
+    attn_out, kv = self.atten_func(x_norm, rope, mask, input_pos, kv_cache)
+    attn_out_norm = self.post_atten_norm(attn_out)
+    x = x + attn_out_norm
+    output = x + self.ff(x)
+    return output, kv
+
+
+class Gemma2(nn.Module):
+  """A Gemma2 model built from the Edge Generative API layers."""
+
+  def __init__(self, config: cfg.ModelConfig):
+    super().__init__()
+
+    # Construct model layers.
+    self.tok_embedding = nn.Embedding(
+        config.vocab_size, config.embedding_dim, padding_idx=0
+    )
+    self.lm_head = nn.Linear(
+        config.embedding_dim,
+        config.vocab_size,
+        bias=config.lm_head_use_bias,
+    )
+    # Gemma2 re-uses the embedding as the head projection layer.
+    self.lm_head.weight.data = self.tok_embedding.weight.data
+    self.transformer_blocks = nn.ModuleList(
+        Gemma2Block(config.block_config(idx), config)
+        for idx in range(config.num_layers)
+    )
+    self.final_norm = builder.build_norm(
+        config.embedding_dim,
+        config.final_norm_config,
+    )
+    # Gemma2 has same hyper parameters for each layer except for attention
+    # types. Use the first layer.
+    attn_config = config.block_config(0).attn_config
+    self.rope_cache = attn_utils.build_rope_cache(
+        size=config.kv_cache_max,
+        dim=int(attn_config.rotary_percentage * attn_config.head_dim),
+        base=10_000,
+        condense_ratio=1,
+        dtype=torch.float32,
+        device=torch.device("cpu"),
+    )
+    self.mask_cache = attn_utils.build_causal_mask_cache(
+        size=config.kv_cache_max,
+        dtype=torch.float32,
+        device=torch.device("cpu"),
+    )
+    self.sliding_window_mask_cache = attn_utils.build_sliding_window_mask_cache(
+        size=config.kv_cache_max,
+        window_size=attn_config.sliding_window_size,
+        dtype=torch.float32,
+        device=torch.device("cpu"),
+    )
+    self.config = config
+
+  def get_attention_mask(
+      self, attn_type: cfg.AttentionType, input_pos: torch.Tensor
+  ) -> torch.Tensor:
+    if attn_type == cfg.AttentionType.LOCAL_SLIDING:
+      return self.sliding_window_mask_cache.index_select(2, input_pos)
+    return self.mask_cache.index_select(2, input_pos)
+
+  @torch.inference_mode
+  def forward(
+      self,
+      tokens: torch.Tensor,
+      input_pos: torch.Tensor,
+      kv_cache: kv_utils.KVCache,
+  ) -> dict[torch.Tensor, kv_utils.KVCache]:
+    _, seq_len = tokens.size()
+    assert self.config.max_seq_len >= seq_len, (
+        f"Cannot forward sequence of length {seq_len}, max seq length is only"
+        f" {self.config.max_seq_len}"
+    )
+    assert len(self.transformer_blocks) == len(kv_cache.caches), (
+        "The number of transformer blocks and the number of KV cache entries"
+        " must be the same."
+    )
+
+    cos, sin = self.rope_cache
+    cos = cos.index_select(0, input_pos)
+    sin = sin.index_select(0, input_pos)
+
+    # token embeddings of shape (b, t, n_embd)
+    x = self.tok_embedding(tokens)
+    x = x * (self.config.embedding_dim**0.5)
+
+    updated_kv_entires = []
+    for i, block in enumerate(self.transformer_blocks):
+      mask = self.get_attention_mask(
+          block.config.attn_config.attn_type, input_pos
+      )
+      kv_entry = kv_cache.caches[i] if kv_cache else None
+      x, kv_entry = block(x, (cos, sin), mask, input_pos, kv_entry)
+      if kv_entry:
+        updated_kv_entires.append(kv_entry)
+    updated_kv_cache = kv_utils.KVCache(tuple(updated_kv_entires))
+
+    x = self.final_norm(x)
+    res = self.lm_head(x)  # (b, t, vocab_size)
+    if self.config.final_logit_softcap is not None:
+      res = res / self.config.final_logit_softcap
+      res = torch.tanh(res)
+      res = res * self.config.final_logit_softcap
+
+    return {"logits": res, "kv_cache": updated_kv_cache}
+
+
+def get_model_config_2b(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for a Gemma2 2B model.
+
+  Args:
+    kv_cache_max_len (int): The maximum sequence length of the KV cache. Default
+      is 1024.
+
+  Returns:
+    The model config for a Gemma 2B model.
+  """
+  norm_config = cfg.NormalizationConfig(
+      type=cfg.NormalizationType.RMS_NORM,
+      epsilon=1e-6,
+      zero_centered=True,
+  )
+  ff_config = cfg.FeedForwardConfig(
+      type=cfg.FeedForwardType.GATED,
+      activation=cfg.ActivationConfig(cfg.ActivationType.GELU_TANH),
+      intermediate_size=9216,
+      pre_ff_norm_config=norm_config,
+      post_ff_norm_config=norm_config,
+  )
+
+  def get_block_config(idx: int) -> cfg.TransformerBlockConfig:
+    attn_config = cfg.AttentionConfig(
+        num_heads=8,
+        head_dim=256,
+        num_query_groups=4,
+        rotary_percentage=1.0,
+        qkv_transpose_before_split=True,
+        logit_softcap=50.0,
+        sliding_window_size=4096,
+        attn_type=(
+            cfg.AttentionType.GLOBAL
+            if idx % 2 == 0
+            else cfg.AttentionType.LOCAL_SLIDING
+        ),
+    )
+    return cfg.TransformerBlockConfig(
+        attn_config=attn_config,
+        ff_config=ff_config,
+        pre_attention_norm_config=norm_config,
+        post_attention_norm_config=norm_config,
+    )
+
+  num_layers = 26
+  config = cfg.ModelConfig(
+      vocab_size=256000,
+      num_layers=num_layers,
+      max_seq_len=8192,
+      embedding_dim=2304,
+      kv_cache_max_len=kv_cache_max_len,
+      block_configs=[get_block_config(i) for i in range(num_layers)],
+      final_norm_config=norm_config,
+      lm_head_use_bias=False,
+      enable_hlfb=True,
+      final_logit_softcap=30.0,
+  )
+  return config
+
+
+def get_fake_model_config(kv_cache_max_len: int = 128) -> cfg.ModelConfig:
+  config = get_model_config_2b(kv_cache_max_len)
+  config.vocab_size = 128
+  config.num_layers = 2
+  config.max_seq_len = 2 * kv_cache_max_len
+  config.embedding_dim = 128
+  config.block_configs = config.block_configs[: config.num_layers]
+  for block_config in config.block_configs:
+    block_config.attn_config.num_heads = 4
+    block_config.attn_config.head_dim = 64
+    block_config.attn_config.sliding_window_size = 64
+    block_config.ff_config.intermediate_size = 128
+  return config
+
+
+def build_2b_model(checkpoint_path: str, **kwargs) -> nn.Module:
+  config = get_model_config_2b(**kwargs)
+  model = Gemma2(config)
+  loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
+  # Since embedding and lm-head use the same weight, we need to set strict
+  # to False.
+  loader.load(model, strict=False)
+  model.eval()
+  return model

ai_edge_torch/generative/examples/gemma/verify_gemma1.py

@@ -0,0 +1,56 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+"""Verifies the reauthored Gemma1 model."""
+
+import logging
+from absl import app
+from absl import flags
+from ai_edge_torch.generative.examples.gemma import gemma1
+from ai_edge_torch.generative.examples.gemma import verify_util
+import kagglehub
+
+
+_PROMPTS = flags.DEFINE_multi_string(
+    "prompts",
+    "What is the meaning of life?",
+    "The input prompts to generate answers.",
+)
+_MAX_NEW_TOKENS = flags.DEFINE_integer(
+    "max_new_tokens",
+    30,
+    "The maximum size of the generated tokens.",
+)
+
+
+def main(_):
+  checkpoint = kagglehub.model_download("google/gemma/pyTorch/2b-it")
+
+  logging.info("Building the reauthored model from: %s", checkpoint)
+  reauthored_model = gemma1.build_2b_model(checkpoint)
+
+  verify_util.verify_reauthored_gemma_model(
+      checkpoint=checkpoint,
+      variant="2b",
+      reauthored_model=reauthored_model,
+      weight_filename="gemma-2b-it.ckpt",
+      generate_prompts=_PROMPTS.value,
+      forward_input_ids=[[1, 2, 3, 4]],
+      max_new_tokens=_MAX_NEW_TOKENS.value,
+  )
+
+
+if __name__ == "__main__":
+  app.run(main)

ai_edge_torch/generative/examples/gemma/verify_gemma2.py

@@ -0,0 +1,57 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+"""Verifies the reauthored Gemma2 model."""
+
+import logging
+from absl import app
+from absl import flags
+from ai_edge_torch.generative.examples.gemma import gemma2
+from ai_edge_torch.generative.examples.gemma import verify_util
+from ai_edge_torch.generative.utilities import verifier
+import kagglehub
+
+
+_PROMPTS = flags.DEFINE_multi_string(
+    "prompts",
+    "What is the meaning of life?",
+    "The input prompts to generate answers.",
+)
+_MAX_NEW_TOKENS = flags.DEFINE_integer(
+    "max_new_tokens",
+    30,
+    "The maximum size of the generated tokens.",
+)
+
+
+def main(_):
+  checkpoint = kagglehub.model_download("google/gemma-2/pyTorch/gemma-2-2b-it")
+
+  logging.info("Building the reauthored model from: %s", checkpoint)
+  reauthored_model = gemma2.build_2b_model(checkpoint)
+
+  verify_util.verify_reauthored_gemma_model(
+      checkpoint=checkpoint,
+      variant="2b-v2",
+      reauthored_model=reauthored_model,
+      generate_prompts=_PROMPTS.value,
+      forward_input_ids=[[2, 651, 9456, 576, 573, 3520, 3858, 603, 235248]],
+      max_new_tokens=_MAX_NEW_TOKENS.value,
+      atol=1e-04,
+  )
+
+
+if __name__ == "__main__":
+  app.run(main)