ai-edge-torch-nightly 0.3.0.dev20240910__py3-none-any.whl → 0.3.0.dev20240913__py3-none-any.whl

ai_edge_torch/generative/examples/gemma/convert_gemma2_to_tflite.py

@@ -13,32 +13,35 @@
 # limitations under the License.
 # ==============================================================================
 
+"""Example of converting a Gemma2 model to multi-signature tflite model."""
+
 import os
-from pathlib import Path
+import pathlib
 
 import ai_edge_torch
 from ai_edge_torch.generative.examples.gemma import gemma2
+from ai_edge_torch.generative.layers import kv_cache as kv_utils
 from ai_edge_torch.generative.quantize import quant_recipes
 import torch
 
 
-def convert_gemma_to_tflite(
+def convert_gemma2_to_tflite(
     checkpoint_path: str,
     prefill_seq_len: int = 512,
     kv_cache_max_len: int = 1024,
     quantize: bool = True,
 ):
-  """Converting a Gemma 2 2B model to multi-signature
-  tflite model.
+  """Converts a Gemma2 2B model to multi-signature tflite model.
 
   Args:
-      checkpoint_path (str): The filepath to the model checkpoint, or directory holding the checkpoint.
+      checkpoint_path (str): The filepath to the model checkpoint, or directory
+        holding the checkpoint.
       prefill_seq_len (int, optional): The maximum size of prefill input tensor.
         Defaults to 512.
       kv_cache_max_len (int, optional): The maximum size of KV cache buffer,
         including both prefill and decode. Defaults to 1024.
-      quantize (bool, optional): Whether the model should be quanized.
-        Defaults to True.
+      quantize (bool, optional): Whether the model should be quanized. Defaults
+        to True.
   """
   pytorch_model = gemma2.build_2b_model(
       checkpoint_path, kv_cache_max_len=kv_cache_max_len
@@ -48,20 +51,36 @@ def convert_gemma_to_tflite(
   prefill_input_pos = torch.arange(0, prefill_seq_len)
   decode_token = torch.tensor([[0]], dtype=torch.long)
   decode_input_pos = torch.tensor([0], dtype=torch.int64)
+  kv = kv_utils.KVCache.from_model_config(pytorch_model.config)
 
   quant_config = quant_recipes.full_int8_dynamic_recipe() if quantize else None
   edge_model = (
       ai_edge_torch.signature(
-          'prefill', pytorch_model, (prefill_tokens, prefill_input_pos)
+          'prefill',
+          pytorch_model,
+          sample_kwargs={
+              'tokens': prefill_tokens,
+              'input_pos': prefill_input_pos,
+              'kv_cache': kv,
+          },
+      )
+      .signature(
+          'decode',
+          pytorch_model,
+          sample_kwargs={
+              'tokens': decode_token,
+              'input_pos': decode_input_pos,
+              'kv_cache': kv,
+          },
       )
-      .signature('decode', pytorch_model, (decode_token, decode_input_pos))
       .convert(quant_config=quant_config)
   )
+  quant_suffix = 'q8' if quantize else 'f32'
   edge_model.export(
-      f'/tmp/gemma2_seq{prefill_seq_len}_kv{kv_cache_max_len}.tflite'
+      f'/tmp/gemma2_{quant_suffix}_seq{prefill_seq_len}_ekv{kv_cache_max_len}.tflite'
   )
 
 
 if __name__ == '__main__':
-  checkpoint_path = os.path.join(Path.home(), 'Downloads/llm_data/gemma2-2b')
-  convert_gemma_to_tflite(checkpoint_path)
+  path = os.path.join(pathlib.Path.home(), 'Downloads/llm_data/gemma2-2b')
+  convert_gemma2_to_tflite(path)

ai_edge_torch/generative/examples/gemma/convert_to_tflite.py

@@ -13,11 +13,14 @@
 # limitations under the License.
 # ==============================================================================
 
+"""Example of converting a Gemma model to multi-signature tflite model."""
+
 import os
-from pathlib import Path
+import pathlib
 
 import ai_edge_torch
 from ai_edge_torch.generative.examples.gemma import gemma
+from ai_edge_torch.generative.layers import kv_cache as kv_utils
 from ai_edge_torch.generative.quantize import quant_recipes
 import torch
 
@@ -48,20 +51,36 @@ def convert_gemma_to_tflite(
   prefill_input_pos = torch.arange(0, prefill_seq_len)
   decode_token = torch.tensor([[0]], dtype=torch.long)
   decode_input_pos = torch.tensor([0], dtype=torch.int64)
+  kv = kv_utils.KVCache.from_model_config(pytorch_model.config)
 
   quant_config = quant_recipes.full_int8_dynamic_recipe() if quantize else None
   edge_model = (
       ai_edge_torch.signature(
-          'prefill', pytorch_model, (prefill_tokens, prefill_input_pos)
+          'prefill',
+          pytorch_model,
+          sample_kwargs={
+              'tokens': prefill_tokens,
+              'input_pos': prefill_input_pos,
+              'kv_cache': kv,
+          },
+      )
+      .signature(
+          'decode',
+          pytorch_model,
+          sample_kwargs={
+              'tokens': decode_token,
+              'input_pos': decode_input_pos,
+              'kv_cache': kv,
+          },
       )
-      .signature('decode', pytorch_model, (decode_token, decode_input_pos))
       .convert(quant_config=quant_config)
   )
+  quant_suffix = 'q8' if quantize else 'f32'
   edge_model.export(
-      f'/tmp/gemma_seq{prefill_seq_len}_kv{kv_cache_max_len}.tflite'
+      f'/tmp/gemma_{quant_suffix}_seq{prefill_seq_len}_ekv{kv_cache_max_len}.tflite'
   )
 
 
 if __name__ == '__main__':
-  checkpoint_path = os.path.join(Path.home(), 'Downloads/llm_data/gemma-2b')
-  convert_gemma_to_tflite(checkpoint_path)
+  path = os.path.join(pathlib.Path.home(), 'Downloads/llm_data/gemma-2b')
+  convert_gemma_to_tflite(path)

ai_edge_torch/generative/examples/gemma/gemma.py

@@ -12,13 +12,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-# Example of building a Gemma model.
+
+"""Example of building a Gemma model."""
 
 import os
-from pathlib import Path
+import pathlib
 
 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
@@ -48,7 +50,6 @@ class Gemma(nn.Module):
   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
@@ -60,18 +61,20 @@ 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(
-        attention.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.attn_config.head_dim
-        ),
+        dim=int(attn_config.rotary_percentage * attn_config.head_dim),
         base=10_000,
         condense_ratio=1,
         dtype=torch.float32,
@@ -84,16 +87,22 @@ class Gemma(nn.Module):
     )
     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:
-    _, seq_len = idx.size()
+  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)
@@ -102,15 +111,20 @@ 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)
 
-    for _, block in enumerate(self.transformer_blocks):
-      x = block(x, (cos, sin), mask, input_pos)
+    updated_kv_entires = []
+    for i, block in enumerate(self.transformer_blocks):
+      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:
@@ -139,18 +153,20 @@ 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,
-      post_attention_norm_config=norm_config,
+      block_configs=block_config,
       final_norm_config=norm_config,
-      parallel_residual=False,
       lm_head_use_bias=False,
       enable_hlfb=True,
   )
@@ -159,7 +175,8 @@ def get_model_config_2b(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
 
 def get_fake_model_config(kv_cache_max_len: int = 128) -> cfg.ModelConfig:
   config = get_model_config_2b(kv_cache_max_len)
-  config.ff_config.intermediate_size = 128
+  # 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
@@ -170,32 +187,35 @@ 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:
+def define_and_run_2b(checkpoint_path: str) -> None:
   """Instantiates and runs a Gemma 2B model."""
 
-  current_dir = Path(__file__).parent.resolve()
+  current_dir = pathlib.Path(__file__).parent.resolve()
   gemma_goldens = torch.load(current_dir / "gemma_lm_logits.pt")
 
   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)
-  lm_logits = model.forward(tokens, input_pos)
+  kv = kv_utils.KVCache.from_model_config(model.config)
+  output = model.forward(tokens, input_pos, kv)
   print("comparing with goldens..")
   assert torch.allclose(
-      gemma_goldens, lm_logits[0, idx.shape[1] - 1, :], atol=1e-05
+      gemma_goldens, output["logits"][0, idx.shape[1] - 1, :], atol=1e-02
   )
 
 
 if __name__ == "__main__":
-  define_and_run_2b()
+  input_checkpoint_path = os.path.join(
+      pathlib.Path.home(), "Downloads/llm_data/gemma-2b"
+  )
+  define_and_run_2b(input_checkpoint_path)

ai_edge_torch/generative/examples/gemma/gemma2.py

@@ -12,14 +12,16 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-# Example of building the Gemma2 2B model.
+
+"""Example of building a Gemma2 model."""
 
 import os
-from pathlib import Path
+import pathlib
 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
@@ -51,7 +53,8 @@ class Gemma2Block(attention.TransformerBlock):
       rope: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
       mask: Optional[torch.Tensor] = None,
       input_pos: Optional[torch.Tensor] = None,
-  ) -> torch.Tensor:
+      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
@@ -62,17 +65,19 @@ class Gemma2Block(attention.TransformerBlock):
       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.
+      output activation from this transformer block, and updated kv cache (if
+      passed in).
     """
 
     x_norm = self.pre_atten_norm(x)
-    attn_out = self.atten_func(x_norm, rope, mask, input_pos)
+    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
+    return output, kv
 
 
 class Gemma2(nn.Module):
@@ -81,7 +86,6 @@ class Gemma2(nn.Module):
   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
@@ -91,20 +95,22 @@ class Gemma2(nn.Module):
         config.vocab_size,
         bias=config.lm_head_use_bias,
     )
-    # Gemma re-uses the embedding as the head projection layer.
+    # 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) for _ in range(config.num_layers)
+        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(
-            config.attn_config.rotary_percentage * config.attn_config.head_dim
-        ),
+        dim=int(attn_config.rotary_percentage * attn_config.head_dim),
         base=10_000,
         condense_ratio=1,
         dtype=torch.float32,
@@ -115,47 +121,56 @@ class Gemma2(nn.Module):
         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=self.config.attn_config.sliding_window_size,
+        window_size=attn_config.sliding_window_size,
         dtype=torch.float32,
         device=torch.device("cpu"),
     )
-
     self.config = config
 
   def get_attention_mask(
-      self, idx: int, input_pos: torch.Tensor
+      self, attn_type: cfg.AttentionType, input_pos: torch.Tensor
   ) -> torch.Tensor:
-    if self.config.attn_config.attn_types:
-      if (
-          self.config.attn_config.attn_types[idx]
-          == cfg.AttentionType.LOCAL_SLIDING
-      ):
-        return self.sliding_window_mask_cache.index_select(2, input_pos)
-
+    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, idx: torch.Tensor, input_pos: torch.Tensor) -> torch.Tensor:
-    _, seq_len = idx.size()
+  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(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):
-      mask = self.get_attention_mask(i, input_pos)
-      x = block(x, (cos, sin), mask, input_pos)
+      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)
@@ -163,7 +178,8 @@ class Gemma2(nn.Module):
       res = res / self.config.final_logit_softcap
       res = torch.tanh(res)
       res = res * self.config.final_logit_softcap
-    return res
+
+    return {"logits": res, "kv_cache": updated_kv_cache}
 
 
 def get_model_config_2b(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
@@ -176,18 +192,6 @@ def get_model_config_2b(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
   Returns:
     The model config for a Gemma 2B model.
   """
-  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_types=[cfg.AttentionType.GLOBAL, cfg.AttentionType.LOCAL_SLIDING]
-      * 13,
-  )
-
   norm_config = cfg.NormalizationConfig(
       type=cfg.NormalizationType.RMS_NORM,
       epsilon=1e-6,
@@ -200,18 +204,38 @@ def get_model_config_2b(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
       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=26,
+      num_layers=num_layers,
       max_seq_len=8192,
       embedding_dim=2304,
       kv_cache_max_len=kv_cache_max_len,
-      attn_config=attn_config,
-      ff_config=ff_config,
-      pre_attention_norm_config=norm_config,
-      post_attention_norm_config=norm_config,
+      block_configs=[get_block_config(i) for i in range(num_layers)],
       final_norm_config=norm_config,
-      parallel_residual=False,
       lm_head_use_bias=False,
       enable_hlfb=True,
       final_logit_softcap=30.0,
@@ -221,14 +245,16 @@ def get_model_config_2b(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
 
 def get_fake_model_config(kv_cache_max_len: int = 128) -> cfg.ModelConfig:
   config = get_model_config_2b(kv_cache_max_len)
-  config.attn_config.num_heads = 4
-  config.attn_config.head_dim = 64
-  config.attn_config.sliding_window_size = 64
-  config.ff_config.intermediate_size = 128
   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
 
 
@@ -236,21 +262,20 @@ 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
+  # 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:
+def define_and_run_2b(checkpoint_path: str) -> None:
   """Instantiates and runs a Gemma2 2B model."""
 
-  current_dir = Path(__file__).parent.resolve()
+  current_dir = pathlib.Path(__file__).parent.resolve()
   gemma2_goldens = torch.load(current_dir / "gemma2it_2b_golden.pt")
   print("Running GEMMA 2")
   kv_cache_max_len = 1024
-  checkpoint_path = os.path.join(Path.home(), "Downloads/llm_data/gemma2-2b")
   model = build_2b_model(checkpoint_path, kv_cache_max_len=kv_cache_max_len)
   toks = torch.from_numpy(
       np.array([2, 651, 9456, 576, 573, 3520, 3858, 603, 235248])
@@ -258,11 +283,13 @@ def define_and_run_2b() -> None:
   tokens = torch.full((1, kv_cache_max_len), 0, dtype=torch.long, device="cpu")
   tokens[0, :9] = toks
   input_pos = torch.arange(0, kv_cache_max_len)
-  out = model.forward(tokens, input_pos)
-  out_final = out[0, 8, :]
+  kv = kv_utils.KVCache.from_model_config(model.config)
+  out = model.forward(tokens, input_pos, kv)
+  out_final = out["logits"][0, 8, :]
   assert torch.allclose(gemma2_goldens, out_final, atol=1e-04)
 
 
 if __name__ == "__main__":
   torch.set_printoptions(sci_mode=True)
-  define_and_run_2b()
+  path = os.path.join(pathlib.Path.home(), "Downloads/llm_data/gemma2-2b")
+  define_and_run_2b(path)

ai_edge_torch/generative/examples/{experimental/phi → phi}/convert_to_tflite.py

@@ -12,16 +12,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-#
-# Note: This is an experimental version of phi2 with external KV cache.
-# Please use with caution.
+
+"""Example of converting a Phi-2 model to multi-signature tflite model."""
 
 import os
-from pathlib import Path
+import pathlib
 
 import ai_edge_torch
-from ai_edge_torch.generative.examples.experimental.phi import phi2
-from ai_edge_torch.generative.layers.experimental import ekv_cache
+from ai_edge_torch.generative.examples.phi import phi2
+from ai_edge_torch.generative.layers import kv_cache
 from ai_edge_torch.generative.quantize import quant_recipes
 import torch
 
@@ -32,9 +31,8 @@ def convert_phi2_to_tflite(
     kv_cache_max_len: int = 1024,
     quantize: bool = True,
 ):
-  """An example method for converting a Phi-2 model to multi-signature
+  """Converts a Phi-2 model to multi-signature tflite model.
 
-  tflite model.
   Args:
       checkpoint_path (str): The filepath to the model checkpoint, or directory
         holding the checkpoint.
@@ -53,7 +51,7 @@ def convert_phi2_to_tflite(
   prefill_input_pos = torch.arange(0, prefill_seq_len)
   decode_token = torch.tensor([[0]], dtype=torch.long)
   decode_input_pos = torch.tensor([0], dtype=torch.int64)
-  kv = ekv_cache.EKVCache.from_model_config(pytorch_model.config)
+  kv = kv_cache.KVCache.from_model_config(pytorch_model.config)
 
   quant_config = quant_recipes.full_int8_dynamic_recipe() if quantize else None
   edge_model = (
@@ -77,11 +75,12 @@ def convert_phi2_to_tflite(
       )
       .convert(quant_config=quant_config)
   )
+  quant_suffix = 'q8' if quantize else 'f32'
   edge_model.export(
-      f'/tmp/phi2_seq{prefill_seq_len}_ekv{kv_cache_max_len}.tflite'
+      f'/tmp/phi2_{quant_suffix}_seq{prefill_seq_len}_ekv{kv_cache_max_len}.tflite'
   )
 
 
 if __name__ == '__main__':
-  checkpoint_path = os.path.join(Path.home(), 'Downloads/llm_data/phi2')
-  convert_phi2_to_tflite(checkpoint_path)
+  path = os.path.join(pathlib.Path.home(), 'Downloads/llm_data/phi2')
+  convert_phi2_to_tflite(path)