ai-edge-torch-nightly 0.3.0.dev20240913__py3-none-any.whl → 0.3.0.dev20240914__py3-none-any.whl

ai_edge_torch/_convert/conversion.py

@@ -17,6 +17,7 @@ import logging
 import os
 from typing import Any, Optional
 
+from ai_edge_torch import fx_pass_base
 from ai_edge_torch import lowertools
 from ai_edge_torch import model
 from ai_edge_torch._convert import fx_passes
@@ -34,7 +35,7 @@ def _run_convert_passes(
   exported_program = generative_fx_passes.run_generative_passes(
       exported_program
   )
-  return fx_passes.run_passes(
+  return fx_pass_base.run_passes(
       exported_program,
       [
           fx_passes.BuildInterpolateCompositePass(),

ai_edge_torch/_convert/fx_passes/__init__.py

@@ -15,44 +15,8 @@
 
 from typing import Sequence, Union
 
-from ai_edge_torch._convert.fx_passes._pass_base import ExportedProgramPassBase
-from ai_edge_torch._convert.fx_passes._pass_base import ExportedProgramPassResult  # NOQA
-from ai_edge_torch._convert.fx_passes._pass_base import FxPassBase
-from ai_edge_torch._convert.fx_passes._pass_base import FxPassResult
-from ai_edge_torch._convert.fx_passes.build_aten_composite_pass import BuildAtenCompositePass  # NOQA
-from ai_edge_torch._convert.fx_passes.build_interpolate_composite_pass import BuildInterpolateCompositePass  # NOQA
-from ai_edge_torch._convert.fx_passes.canonicalize_pass import CanonicalizePass
-from ai_edge_torch._convert.fx_passes.inject_mlir_debuginfo_pass import InjectMlirDebuginfoPass  # NOQA
-from ai_edge_torch._convert.fx_passes.optimize_layout_transposes_pass import OptimizeLayoutTransposesPass  # NOQA
-from torch.export import ExportedProgram
-from torch.fx.passes.infra.pass_manager import pass_result_wrapper
-import torch.utils._pytree as pytree
-
-
-# TODO(cnchan): make a PassManager class.
-def run_passes(
-    exported_program: ExportedProgram,
-    passes: Sequence[Union[ExportedProgramPassBase, FxPassBase]],
-) -> ExportedProgram:
-  passes, _ = pytree.tree_flatten(passes)
-  for pass_ in passes:
-    if not isinstance(pass_, ExportedProgramPassBase):
-      pass_ = pass_result_wrapper(pass_)
-    if isinstance(pass_, ExportedProgramPassBase):
-      exported_program = pass_(exported_program).exported_program
-    else:
-      gm = exported_program.graph_module
-      gm, modified = pass_(gm)
-      if modified and gm is not exported_program.graph_module:
-        exported_program = ExportedProgram(
-            root=gm,
-            graph=gm.graph,
-            graph_signature=exported_program.graph_signature,
-            state_dict=exported_program.state_dict,
-            range_constraints=exported_program.range_constraints,
-            module_call_graph=exported_program.module_call_graph,
-            example_inputs=exported_program.example_inputs,
-            verifier=exported_program.verifier,
-            constants=exported_program.constants,
-        )
-  return exported_program
+from ai_edge_torch._convert.fx_passes.build_aten_composite_pass import BuildAtenCompositePass
+from ai_edge_torch._convert.fx_passes.build_interpolate_composite_pass import BuildInterpolateCompositePass
+from ai_edge_torch._convert.fx_passes.inject_mlir_debuginfo_pass import InjectMlirDebuginfoPass
+from ai_edge_torch._convert.fx_passes.optimize_layout_transposes_pass import OptimizeLayoutTransposesPass
+from ai_edge_torch.fx_pass_base import CanonicalizePass

ai_edge_torch/_convert/fx_passes/build_aten_composite_pass.py

@@ -13,11 +13,10 @@
 # limitations under the License.
 # ==============================================================================
 
-from functools import reduce
 from typing import Any, Callable
+from ai_edge_torch import fx_pass_base
 from ai_edge_torch import lowertools
 import torch
-from torch.fx.passes.infra import pass_base
 import torch.utils._pytree as pytree
 
 _composite_builders: dict[
@@ -277,7 +276,7 @@ def _aten_embedding(gm: torch.fx.GraphModule, node: torch.fx.Node):
   node.target = embedding
 
 
-class BuildAtenCompositePass(pass_base.PassBase):
+class BuildAtenCompositePass(fx_pass_base.PassBase):
 
   def call(self, graph_module: torch.fx.GraphModule):
     for node in graph_module.graph.nodes:
@@ -286,4 +285,4 @@ class BuildAtenCompositePass(pass_base.PassBase):
 
     graph_module.graph.lint()
     graph_module.recompile()
-    return pass_base.PassResult(graph_module, True)
+    return fx_pass_base.PassResult(graph_module, True)

ai_edge_torch/_convert/fx_passes/build_interpolate_composite_pass.py

@@ -16,8 +16,7 @@
 
 import functools
 
-from ai_edge_torch._convert.fx_passes._pass_base import ExportedProgramPassBase
-from ai_edge_torch._convert.fx_passes._pass_base import ExportedProgramPassResult  # NOQA
+from ai_edge_torch import fx_pass_base
 from ai_edge_torch.hlfb import mark_pattern
 from ai_edge_torch.hlfb.mark_pattern import pattern as pattern_module
 import torch
@@ -103,7 +102,7 @@ def _get_interpolate_nearest2d_pattern():
   return pattern
 
 
-class BuildInterpolateCompositePass(ExportedProgramPassBase):
+class BuildInterpolateCompositePass(fx_pass_base.ExportedProgramPassBase):
 
   def __init__(self):
     super().__init__()
@@ -124,4 +123,4 @@ class BuildInterpolateCompositePass(ExportedProgramPassBase):
 
     graph_module.graph.lint()
     graph_module.recompile()
-    return ExportedProgramPassResult(exported_program, True)
+    return fx_pass_base.ExportedProgramPassResult(exported_program, True)

ai_edge_torch/_convert/fx_passes/inject_mlir_debuginfo_pass.py

@@ -13,10 +13,9 @@
 # limitations under the License.
 # ==============================================================================
 
+from ai_edge_torch import fx_pass_base
 from ai_edge_torch import lowertools
 import torch
-from torch.fx.passes.infra.pass_base import PassBase
-from torch.fx.passes.infra.pass_base import PassResult
 import torch.utils._pytree as pytree
 
 
@@ -62,7 +61,7 @@ def _wrap_call_function_node_with_debuginfo_writer(node: torch.fx.GraphModule):
   node.target = debuginfo_writer
 
 
-class InjectMlirDebuginfoPass(PassBase):
+class InjectMlirDebuginfoPass(fx_pass_base.PassBase):
 
   def call(self, graph_module: torch.fx.GraphModule):
     for node in graph_module.graph.nodes:
@@ -70,4 +69,4 @@ class InjectMlirDebuginfoPass(PassBase):
 
     graph_module.graph.lint()
     graph_module.recompile()
-    return PassResult(graph_module, True)
+    return fx_pass_base.PassResult(graph_module, True)

ai_edge_torch/_convert/fx_passes/optimize_layout_transposes_pass/pass_body.py

@@ -18,8 +18,7 @@ import operator
 import os
 from typing import Union
 
-from ai_edge_torch._convert.fx_passes import ExportedProgramPassBase
-from ai_edge_torch._convert.fx_passes import ExportedProgramPassResult
+from ai_edge_torch import fx_pass_base
 from ai_edge_torch._convert.fx_passes.optimize_layout_transposes_pass import layout_check  # NOQA
 from ai_edge_torch._convert.fx_passes.optimize_layout_transposes_pass import layout_mark  # NOQA
 from ai_edge_torch._convert.fx_passes.optimize_layout_transposes_pass import layout_partitioners  # NOQA
@@ -31,7 +30,7 @@ import torch.ao.quantization.quantize_pt2e
 TransposeFunc = Union[utils.tensor_to_nchw, utils.tensor_to_nhwc]
 
 
-class OptimizeLayoutTransposesPass(ExportedProgramPassBase):
+class OptimizeLayoutTransposesPass(fx_pass_base.ExportedProgramPassBase):
 
   def get_source_meta(self, node: torch.fx.Node):
     keys = ["stack_trace", "nn_module_stack", "source_fn_stack", "from_node"]
@@ -94,7 +93,7 @@ class OptimizeLayoutTransposesPass(ExportedProgramPassBase):
 
     q_args = input_q.args[1:]
     q_kwargs = input_q.kwargs
-    q_op, dq_op = self.get_paired_q_dq_ops(input_q.target)
+    q_op, dq_op = utils.get_paired_q_dq_ops(input_q.target)
     with graph.inserting_before(target):
       # Q and DQ inserted here may required updating the `axis` arg when they
       # are per_channel ops. However, instead of updating here, the nodes would
@@ -301,4 +300,4 @@ class OptimizeLayoutTransposesPass(ExportedProgramPassBase):
     # Mark const node again for debugging
     self.mark_const_nodes(exported_program)
 
-    return ExportedProgramPassResult(exported_program, True)
+    return fx_pass_base.ExportedProgramPassResult(exported_program, True)

ai_edge_torch/config.py

@@ -21,4 +21,7 @@ import os
 
 @dataclasses.dataclass
 class Config:
-  use_torch_xla: bool = os.environ.get("USE_TORCH_XLA", "True") == "True"
+  use_torch_xla: bool = os.environ.get("USE_TORCH_XLA", "true").lower() in (
+      "1",
+      "true",
+  )

ai_edge_torch/fx_pass_base.py

@@ -0,0 +1,101 @@
+# 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.
+# ==============================================================================
+
+import abc
+import collections
+from typing import Sequence, Union
+
+import torch
+from torch.fx.passes.infra.pass_base import PassBase
+from torch.fx.passes.infra.pass_base import PassResult
+from torch.fx.passes.infra.pass_manager import pass_result_wrapper
+import torch.utils._pytree as pytree
+
+FxPassBase = PassBase
+FxPassResult = PassResult
+ExportedProgramPassResult = collections.namedtuple(
+    "ExportedProgramPassResult", ["exported_program", "modified"]
+)
+
+
+class ExportedProgramPassBase(abc.ABC):
+
+  def __call__(
+      self, exported_program: torch.export.ExportedProgram
+  ) -> ExportedProgramPassResult:
+    self.requires(exported_program)
+    res = self.call(exported_program)
+    self.ensures(exported_program)
+    return res
+
+  @abc.abstractmethod
+  def call(
+      self, exported_program: torch.export.ExportedProgram
+  ) -> ExportedProgramPassResult:
+    pass
+
+  def requires(self, exported_program: torch.export.ExportedProgram) -> None:
+    pass
+
+  def ensures(self, exported_program: torch.export.ExportedProgram) -> None:
+    pass
+
+
+# TODO(cnchan): make a PassManager class.
+def run_passes(
+    exported_program: torch.export.ExportedProgram,
+    passes: Sequence[Union[ExportedProgramPassBase, FxPassBase]],
+) -> torch.export.ExportedProgram:
+  passes, _ = pytree.tree_flatten(passes)
+  for pass_ in passes:
+    if not isinstance(pass_, ExportedProgramPassBase):
+      pass_ = pass_result_wrapper(pass_)
+    if isinstance(pass_, ExportedProgramPassBase):
+      exported_program = pass_(exported_program).exported_program
+    else:
+      gm = exported_program.graph_module
+      gm, modified = pass_(gm)
+      if modified and gm is not exported_program.graph_module:
+        exported_program = torch.export.ExportedProgram(
+            root=gm,
+            graph=gm.graph,
+            graph_signature=exported_program.graph_signature,
+            state_dict=exported_program.state_dict,
+            range_constraints=exported_program.range_constraints,
+            module_call_graph=exported_program.module_call_graph,
+            example_inputs=exported_program.example_inputs,
+            verifier=exported_program.verifier,
+            constants=exported_program.constants,
+        )
+  return exported_program
+
+
+class CanonicalizePass(ExportedProgramPassBase):
+
+  # A dummy decomp table for running ExportedProgram.run_decompositions without
+  # any op decompositions but just aot_export_module. Due to the check in
+  # run_decompositions, if None or an empty dict is passed as decomp_table,
+  # it will run the default aten-coreaten decompositions. Therefore a non-empty
+  # dummy decomp table is needed.
+  # Ref: https://github.com/pytorch/pytorch/blob/db895ace1d36726e64781774f53b3d3098206116/torch/export/exported_program.py#L543
+  _DUMMY_DECOMP_TABLE = {
+      torch._ops.OperatorBase(): lambda: None,
+  }
+
+  def call(self, exported_program: torch.export.ExportedProgram):
+    exported_program = exported_program.run_decompositions(
+        self._DUMMY_DECOMP_TABLE
+    )
+    return ExportedProgramPassResult(exported_program, True)

ai_edge_torch/generative/examples/gemma/convert_gemma2_to_tflite.py

@@ -47,10 +47,10 @@ def convert_gemma2_to_tflite(
       checkpoint_path, kv_cache_max_len=kv_cache_max_len
   )
   # Tensors used to trace the model graph during conversion.
-  prefill_tokens = torch.full((1, prefill_seq_len), 0, dtype=torch.long)
-  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)
+  prefill_tokens = torch.full((1, prefill_seq_len), 0, dtype=torch.int)
+  prefill_input_pos = torch.arange(0, prefill_seq_len, dtype=torch.int)
+  decode_token = torch.tensor([[0]], dtype=torch.int)
+  decode_input_pos = torch.tensor([0], dtype=torch.int)
   kv = kv_utils.KVCache.from_model_config(pytorch_model.config)
 
   quant_config = quant_recipes.full_int8_dynamic_recipe() if quantize else None

ai_edge_torch/generative/examples/gemma/convert_to_tflite.py

@@ -47,10 +47,10 @@ def convert_gemma_to_tflite(
       checkpoint_path, kv_cache_max_len=kv_cache_max_len
   )
   # Tensors used to trace the model graph during conversion.
-  prefill_tokens = torch.full((1, prefill_seq_len), 0, dtype=torch.long)
-  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)
+  prefill_tokens = torch.full((1, prefill_seq_len), 0, dtype=torch.int)
+  prefill_input_pos = torch.arange(0, prefill_seq_len, dtype=torch.int)
+  decode_token = torch.tensor([[0]], dtype=torch.int)
+  decode_input_pos = torch.tensor([0], dtype=torch.int)
   kv = kv_utils.KVCache.from_model_config(pytorch_model.config)
 
   quant_config = quant_recipes.full_int8_dynamic_recipe() if quantize else None

ai_edge_torch/generative/examples/gemma/gemma.py

@@ -203,9 +203,9 @@ def define_and_run_2b(checkpoint_path: str) -> None:
   kv_cache_max_len = 1024
   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 = torch.full((1, kv_cache_max_len), 0, dtype=torch.int, device="cpu")
   tokens[0, :4] = idx
-  input_pos = torch.arange(0, kv_cache_max_len)
+  input_pos = torch.arange(0, kv_cache_max_len, dtype=torch.int)
   kv = kv_utils.KVCache.from_model_config(model.config)
   output = model.forward(tokens, input_pos, kv)
   print("comparing with goldens..")

ai_edge_torch/generative/examples/gemma/gemma2.py

@@ -280,9 +280,9 @@ def define_and_run_2b(checkpoint_path: str) -> None:
   toks = torch.from_numpy(
       np.array([2, 651, 9456, 576, 573, 3520, 3858, 603, 235248])
   )
-  tokens = torch.full((1, kv_cache_max_len), 0, dtype=torch.long, device="cpu")
+  tokens = torch.full((1, kv_cache_max_len), 0, dtype=torch.int, device="cpu")
   tokens[0, :9] = toks
-  input_pos = torch.arange(0, kv_cache_max_len)
+  input_pos = torch.arange(0, kv_cache_max_len, dtype=torch.int)
   kv = kv_utils.KVCache.from_model_config(model.config)
   out = model.forward(tokens, input_pos, kv)
   out_final = out["logits"][0, 8, :]

ai_edge_torch/generative/examples/openelm/convert_to_tflite.py

@@ -0,0 +1,86 @@
+# 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 converting OpenELM model to multi-signature tflite model."""
+
+import os
+import pathlib
+
+import ai_edge_torch
+from ai_edge_torch.generative.examples.openelm import openelm
+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_openelm_to_tflite(
+    checkpoint_path: str,
+    prefill_seq_len: int = 512,
+    kv_cache_max_len: int = 1024,
+    quantize: bool = True,
+):
+  """Converts OpenELM model to multi-signature tflite model.
+
+  Args:
+      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.
+  """
+  pytorch_model = openelm.build_model(
+      checkpoint_path, kv_cache_max_len=kv_cache_max_len
+  )
+  # Tensors used to trace the model graph during conversion.
+  prefill_tokens = torch.full((1, prefill_seq_len), 0, dtype=torch.int)
+  prefill_input_pos = torch.arange(0, prefill_seq_len, dtype=torch.int)
+  decode_token = torch.tensor([[0]], dtype=torch.int)
+  decode_input_pos = torch.tensor([0], dtype=torch.int)
+  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,
+          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,
+          },
+      )
+      .convert(quant_config=quant_config)
+  )
+  quant_suffix = 'q8' if quantize else 'f32'
+  edge_model.export(
+      f'/tmp/openelm_{quant_suffix}_seq{prefill_seq_len}_ekv{kv_cache_max_len}.tflite'
+  )
+
+
+if __name__ == '__main__':
+  path = os.path.join(pathlib.Path.home(), 'Downloads/llm_data/openelm')
+  convert_openelm_to_tflite(path)

ai_edge_torch/generative/examples/openelm/openelm.py

@@ -0,0 +1,237 @@
+# 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 an OpenELM model."""
+
+import os
+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
+import numpy as np
+import torch
+from torch import nn
+
+TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
+    ff_up_proj="transformer.layers.{}.ffn.proj_1",
+    ff_down_proj="transformer.layers.{}.ffn.proj_2",
+    attn_fused_qkv_proj="transformer.layers.{}.attn.qkv_proj",
+    attn_query_norm="transformer.layers.{}.attn.q_norm",
+    attn_key_norm="transformer.layers.{}.attn.k_norm",
+    attn_output_proj="transformer.layers.{}.attn.out_proj",
+    pre_attn_norm="transformer.layers.{}.attn_norm",
+    pre_ff_norm="transformer.layers.{}.ffn_norm",
+    embedding="transformer.token_embeddings",
+    final_norm="transformer.norm",
+    lm_head=None,
+)
+
+
+class OpenELM(nn.Module):
+  """An OpenELM 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
+    )
+    # OpenELM re-uses the embedding as the head projection layer.
+    self.lm_head.weight.data = self.tok_embedding.weight.data
+    self.transformer_blocks = nn.ModuleList(
+        attention.TransformerBlock(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,
+    )
+    # OpenELM has same hyper parameters for rotary_percentage and head_dim for
+    # each layer block. Use the first block.
+    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.config = config
+
+  @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)
+    mask = self.mask_cache.index_select(2, input_pos)
+    mask = mask[:, :, :, : self.config.kv_cache_max]
+
+    # token embeddings of shape (b, t, n_embd)
+    x = self.tok_embedding(tokens)
+
+    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)
+    logits = self.lm_head(x)  # (b, t, vocab_size)
+    return {"logits": logits, "kv_cache": updated_kv_cache}
+
+
+def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  """Returns the model config for an OpenELM model.
+
+  Args:
+    kv_cache_max_len (int): The maximum sequence length of the KV cache. Default
+      is 1024.
+
+  Returns:
+    The model config for an OpenELM model.
+  """
+  norm_config = cfg.NormalizationConfig(
+      type=cfg.NormalizationType.RMS_NORM, epsilon=1e-6
+  )
+  num_heads = [12] * 4 + [16] * 14 + [20] * 12 + [24] * 6
+  num_query_groups = [3] * 4 + [4] * 14 + [5] * 12 + [6] * 6
+
+  def make_divisible(v, d):
+    """Ensures that all layers have a channel number that is divisible by d."""
+    new_v = int(v + d / 2) // d * d
+    # Make sure that round down does not go down by more than 10%.
+    if new_v < 0.9 * v:
+      new_v += d
+    return new_v
+
+  # The way to get intermediate size is from
+  # https://huggingface.co/apple/OpenELM-3B/blob/main/modeling_openelm.py
+  def get_intermediate_size(idx: int) -> int:
+    return make_divisible((0.5 + 0.1 * idx) * 3072, 256)
+
+  def get_block_config(idx: int) -> cfg.TransformerBlockConfig:
+    return cfg.TransformerBlockConfig(
+        attn_config=cfg.AttentionConfig(
+            num_heads=num_heads[idx],
+            head_dim=128,
+            num_query_groups=num_query_groups[idx],
+            rotary_percentage=1.0,
+            qkv_transpose_before_split=True,
+            query_norm_config=norm_config,
+            key_norm_config=norm_config,
+        ),
+        ff_config=cfg.FeedForwardConfig(
+            type=cfg.FeedForwardType.SEQUENTIAL,
+            activation=cfg.ActivationConfig(
+                cfg.ActivationType.SILU_GLU, gate_is_front=True
+            ),
+            intermediate_size=get_intermediate_size(idx),
+            pre_ff_norm_config=norm_config,
+        ),
+        pre_attention_norm_config=norm_config,
+    )
+
+  num_layers = 36
+  config = cfg.ModelConfig(
+      vocab_size=32000,
+      num_layers=num_layers,
+      max_seq_len=2048,
+      embedding_dim=3072,
+      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,
+  )
+  return config
+
+
+def get_fake_model_config(kv_cache_max_len: int = 128) -> cfg.ModelConfig:
+  config = get_model_config(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 = 3
+    block_config.attn_config.head_dim = 64
+    block_config.ff_config.intermediate_size = 128
+  return config
+
+
+def build_model(checkpoint_path: str, **kwargs) -> nn.Module:
+  config = get_model_config(**kwargs)
+  model = OpenELM(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
+
+
+def define_and_run(checkpoint_path: str) -> None:
+  """Instantiates and runs an OpenELM model."""
+
+  current_dir = pathlib.Path(__file__).parent.resolve()
+  openelm_goldens = torch.load(current_dir / "openelm_lm_logits.pt")
+  kv_cache_max_len = 1024
+  model = build_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.int, device="cpu")
+  tokens[0, :4] = idx
+  input_pos = torch.arange(0, kv_cache_max_len, dtype=torch.int)
+  kv = kv_utils.KVCache.from_model_config(model.config)
+  output = model.forward(tokens, input_pos, kv)
+  assert torch.allclose(
+      openelm_goldens, output["logits"][0, idx.shape[1] - 1, :], atol=1e-05
+  )
+
+
+if __name__ == "__main__":
+  input_checkpoint_path = os.path.join(
+      pathlib.Path.home(), "Downloads/llm_data/openelm"
+  )
+  define_and_run(input_checkpoint_path)