ai-edge-torch-nightly 0.2.0.dev20240718__py3-none-any.whl → 0.2.0.dev20240720__py3-none-any.whl

ai_edge_torch/generative/examples/tiny_llama/tiny_llama.py

@@ -149,6 +149,8 @@ def build_model(checkpoint_path, **kwargs) -> nn.Module:
 
 
 def define_and_run() -> None:
+  current_dir = Path(__file__).parent.resolve()
+  tiny_llama_goldens = torch.load(current_dir / "tiny_llama_lm_logits.pt")
   kv_cache_max_len = 1024
   checkpoint_path = os.path.join(Path.home(), "Downloads/llm_data/tiny_llama")
   model = build_model(checkpoint_path, kv_cache_max_len=kv_cache_max_len)
@@ -156,8 +158,10 @@ def define_and_run() -> None:
   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))
+  lm_logits = model.forward(tokens, input_pos)
+  assert torch.allclose(
+      tiny_llama_goldens, lm_logits[0, idx.shape[1] - 1, :], atol=1e-05
+  )
 
 
 if __name__ == "__main__":

ai_edge_torch/generative/test/test_experimental_ekv.py

@@ -0,0 +1,122 @@
+# 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.
+# ==============================================================================
+# A suite of tests to validate experimental external KV Cache layers and models.
+
+import unittest
+
+import numpy as np
+import torch
+
+from ai_edge_torch.generative.examples.experimental.gemma import gemma
+from ai_edge_torch.generative.examples.experimental.phi import phi2
+from ai_edge_torch.generative.examples.experimental.tiny_llama import tiny_llama  # NOQA
+from ai_edge_torch.generative.layers.experimental import ekv_cache as kv_utils
+import ai_edge_torch.generative.layers.model_config as cfg
+
+
+class TestExternalKVLayers(unittest.TestCase):
+
+  def _get_test_config(self, num_layers, head_dim, num_query_groups, kv_cache_max_len):
+    attn_config = cfg.AttentionConfig(num_heads=1, num_query_groups=num_query_groups)
+    config = cfg.ModelConfig(
+        kv_cache_max_len=kv_cache_max_len,
+        embedding_dim=head_dim,
+        attn_config=attn_config,
+        num_layers=num_layers,
+        max_seq_len=None,
+        vocab_size=None,
+        ff_config=None,
+    )
+    return config
+
+  def test_cache_udpate(self):
+    N = 1
+    HEAD_DIM = 2
+    NUM_QG = 1
+    KV_LEN = 4
+    config = self._get_test_config(
+        num_layers=N,
+        head_dim=HEAD_DIM,
+        num_query_groups=NUM_QG,
+        kv_cache_max_len=KV_LEN,
+    )
+    kv = kv_utils.EKVCache.from_model_config(config)
+    entry = kv.caches[0]
+    # single-slice update
+    input_pos = torch.tensor([1])
+    k_slice = v_slice = torch.full((1, 1, NUM_QG, HEAD_DIM), 5, dtype=torch.float)
+    updated_entry = kv_utils.update(entry, input_pos, k_slice, v_slice)
+    self.assertEqual(
+        updated_entry.k_cache.numpy().flatten().tolist(), [0, 0, 5, 5, 0, 0, 0, 0]
+    )
+    self.assertEqual(
+        updated_entry.v_cache.numpy().flatten().tolist(), [0, 0, 5, 5, 0, 0, 0, 0]
+    )
+    # multi-slice update
+    input_pos = torch.tensor([0, 3])
+    k_slice = v_slice = torch.full((1, 2, NUM_QG, HEAD_DIM), 7, dtype=torch.float)
+    updated_entry = kv_utils.update(entry, input_pos, k_slice, v_slice)
+    self.assertEqual(
+        updated_entry.k_cache.numpy().flatten().tolist(), [7, 7, 0, 0, 0, 0, 7, 7]
+    )
+    self.assertEqual(
+        updated_entry.v_cache.numpy().flatten().tolist(), [7, 7, 0, 0, 0, 0, 7, 7]
+    )
+
+  def test_serialization(self):
+    class TestModel(torch.nn.Module):
+
+      def forward(self, kv: kv_utils.EKVCache) -> kv_utils.EKVCache:
+        updated_kv_entries = [
+            kv_utils.KVCacheEntry(
+                torch.zeros_like(entry.k_cache), torch.zeros_like(entry.v_cache)
+            )
+            for entry in kv.caches
+        ]
+        return kv_utils.EKVCache(updated_kv_entries)
+
+    N = 1
+    HEAD_DIM = 2
+    NUM_QG = 1
+    KV_LEN = 4
+    config = self._get_test_config(
+        num_layers=N,
+        head_dim=HEAD_DIM,
+        num_query_groups=NUM_QG,
+        kv_cache_max_len=KV_LEN,
+    )
+    kv = kv_utils.EKVCache.from_model_config(config)
+    model = TestModel()
+    exported_program = torch.export.export(model, (kv,))
+    input_specs = exported_program.graph_signature.input_specs
+    self.assertEqual(len(input_specs), 2)
+    self.assertEqual(input_specs[0].arg.name, "kv_k_0")
+    self.assertEqual(input_specs[1].arg.name, "kv_v_0")
+
+
+class TestExternalKVModels(unittest.TestCase):
+
+  def test_can_build_gemma(self):
+    gemma.define_and_run_2b(checkpoint_path=None, test_model=True)
+
+  def test_can_build_phi2(self):
+    phi2.define_and_run(checkpoint_path=None, test_model=True)
+
+  def test_can_build_tinyllama(self):
+    tiny_llama.define_and_run(checkpoint_path=None, test_model=True)
+
+
+if __name__ == "__main__":
+  unittest.main()

{ai_edge_torch_nightly-0.2.0.dev20240718.dist-info → ai_edge_torch_nightly-0.2.0.dev20240720.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ai-edge-torch-nightly
-Version: 0.2.0.dev20240718
+Version: 0.2.0.dev20240720
 Summary: Supporting PyTorch models with the Google AI Edge TFLite runtime.
 Home-page: https://github.com/google-ai-edge/ai-edge-torch
 Keywords: On-Device ML,AI,Google,TFLite,PyTorch,LLMs,GenAI

{ai_edge_torch_nightly-0.2.0.dev20240718.dist-info → ai_edge_torch_nightly-0.2.0.dev20240720.dist-info}/RECORD

@@ -2,7 +2,7 @@ ai_edge_torch/__init__.py,sha256=CNDboRP4zQBpz2hznNCQWcQCARvNXUm3DMa1Dw_XXFg,106
 ai_edge_torch/model.py,sha256=8Ba9ia7TCM_fciulw6qObmzdcxL3IaLQKDqpR7Lxp-Q,4440
 ai_edge_torch/convert/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/convert/conversion.py,sha256=StJHglvx6cii36oi8sj-tZda009e9UqR6ufZOZkP1SY,4137
-ai_edge_torch/convert/conversion_utils.py,sha256=PKXIlSCU-8DhppNBh9ICDNUlEOpV0HgCbt85jDVe3rA,13394
+ai_edge_torch/convert/conversion_utils.py,sha256=TA-fbRApU_wdZYg8VmQSGiH4sm70iITsLRBBi5vODTw,13813
 ai_edge_torch/convert/converter.py,sha256=hSrW6A-kix9cjdD6CuLL7rseWrLKoV6GRy-iUSW_nZc,7875
 ai_edge_torch/convert/to_channel_last_io.py,sha256=zo5tY3yDhY_EPCkrL1XSXs2uRFS8B4_qu08dSjNsUGk,2778
 ai_edge_torch/convert/fx_passes/__init__.py,sha256=EPs4PSIDLuRH5EBETi6deaOvaaf_Q4xD3_9NVcR7x8o,2810
@@ -22,7 +22,7 @@ ai_edge_torch/convert/fx_passes/optimize_layout_transposes_pass/layout_partition
 ai_edge_torch/convert/fx_passes/optimize_layout_transposes_pass/layout_partitioners/greedy.py,sha256=8uHJbIwPMTgeSfYVba163pkXSQkHLxFwar_8A1AhgAM,2279
 ai_edge_torch/convert/fx_passes/optimize_layout_transposes_pass/layout_partitioners/min_cut.py,sha256=lklGxE1R32vsjFbhLLBDEFL4pfLi_iTgI9Ftb6Grezk,7156
 ai_edge_torch/convert/test/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
-ai_edge_torch/convert/test/test_convert.py,sha256=h0vOffr8saDQRkiXljNWDZ17EBjnS4xAtxd8DxETleY,9081
+ai_edge_torch/convert/test/test_convert.py,sha256=itOZDKsh0-0aoly1b1M72M179Yr2BJqtTe6ivueZSc4,12607
 ai_edge_torch/convert/test/test_convert_composites.py,sha256=8UkdPtGkjgSVLCzB_rpM2FmwYuMyt6WE48umX_kr_Sg,7601
 ai_edge_torch/convert/test/test_convert_multisig.py,sha256=kMaGnHe9ylfyU68qCifYcaGwJqyejKz--QQt9jS2oUA,4537
 ai_edge_torch/convert/test/test_to_channel_last_io.py,sha256=I8c4ZG3v1vo0yxQYzLK_BTId4AOL9vadHGDtfCUZ4UI,2930
@@ -35,12 +35,22 @@ ai_edge_torch/debug/test/test_search_model.py,sha256=0guAEon5cvwBpPXk6J0wVOKj7TX
 ai_edge_torch/experimental/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
+ai_edge_torch/generative/examples/experimental/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
+ai_edge_torch/generative/examples/experimental/gemma/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
+ai_edge_torch/generative/examples/experimental/gemma/convert_to_tflite.py,sha256=bW_KOj_3fcZAggfST3zHWcMcNJs70b0pld-vvauAOgo,3076
+ai_edge_torch/generative/examples/experimental/gemma/gemma.py,sha256=u4DNsZRnN7whDoK8yQet9Yahb01ToVqTuFQmWV1__1g,6606
+ai_edge_torch/generative/examples/experimental/phi/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
+ai_edge_torch/generative/examples/experimental/phi/convert_to_tflite.py,sha256=sLU35tpQ-PEbhZbLfC1vSqM-HamKREVBpIoywWh9O3M,3036
+ai_edge_torch/generative/examples/experimental/phi/phi2.py,sha256=zgxB2JSFAevyS28C6-wIBaQeeKTUejUJY4dnR4BqRBI,6150
+ai_edge_torch/generative/examples/experimental/tiny_llama/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
+ai_edge_torch/generative/examples/experimental/tiny_llama/convert_to_tflite.py,sha256=PChEhBotZ8k6GZiq9e_AYnn3RyhNIVm_U96QhVjx3jY,3126
+ai_edge_torch/generative/examples/experimental/tiny_llama/tiny_llama.py,sha256=1vL0u6Pkd8SV8uei9BGzSAIokclT_RaE3K0IczoPfeI,6291
 ai_edge_torch/generative/examples/gemma/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/gemma/convert_to_tflite.py,sha256=UMEZGDGhFvAX4eT5KHAE1Xbxw-qtQWEMxgvB8cSH6wY,2531
-ai_edge_torch/generative/examples/gemma/gemma.py,sha256=1lZfXGHmbII4rFu0U2B9NzlJCRhphxtmQtkCHQ39_uw,5935
+ai_edge_torch/generative/examples/gemma/gemma.py,sha256=YyGGsgEByIg_tIysMBqaBztf_csthZIjah8mmH5o7UA,6144
 ai_edge_torch/generative/examples/phi2/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/phi2/convert_to_tflite.py,sha256=uF1A2EX8xYie30-T2Z7s1WZCtFhp5CEwRV8SCd7Umrc,2505
-ai_edge_torch/generative/examples/phi2/phi2.py,sha256=PMhKC6JCAMYSj2F3UmWHWK4rTcXD-B6PuehaoDccRqk,5562
+ai_edge_torch/generative/examples/phi2/phi2.py,sha256=KjfTrD2OBzOfq83-XvJ6ZhmXLuP_VqugSOwyj-M5YY4,5767
 ai_edge_torch/generative/examples/stable_diffusion/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/stable_diffusion/attention.py,sha256=Lo4Dq7a3Kg-lyH56iqGtqCo5UaClQHRCTDdNagXGTo8,3535
 ai_edge_torch/generative/examples/stable_diffusion/clip.py,sha256=P-cUUQaQKGKV2p-7hvLJ--RpCIA7gk8WCDRgg0pNtd0,4331
@@ -58,15 +68,15 @@ ai_edge_torch/generative/examples/stable_diffusion/samplers/k_lms.py,sha256=iPYX
 ai_edge_torch/generative/examples/stable_diffusion/samplers/sampler.py,sha256=5iRfU5MO6GR6K3WrdddIU_9U7ZZGEEb7zGKVY1WFl-8,1340
 ai_edge_torch/generative/examples/t5/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/t5/convert_to_tflite.py,sha256=7RwaZQaKhFt3zKAUbFjq95CSYhL1nd9BVSbSRNJp4-4,4529
-ai_edge_torch/generative/examples/t5/t5.py,sha256=L6YrVzUEzP-Imb8W28LdukFGrx1aWSzz1kyYK_9RFZM,21087
+ai_edge_torch/generative/examples/t5/t5.py,sha256=fVtJ0S8v2bMtvEuDqD6Orw7CTyXqnRIqZfKcz7DBeJc,21212
 ai_edge_torch/generative/examples/t5/t5_attention.py,sha256=KaGzCAViNOpJIQbRF-ItouuVPqI9nroWRRGN-KFYKZs,8357
 ai_edge_torch/generative/examples/test_models/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/examples/test_models/toy_model.py,sha256=Sf3ZMYv-iuMRKAKLow47qth8vTF1zl6i8TxJ9uT_StU,3885
-ai_edge_torch/generative/examples/test_models/toy_model_with_external_kv_cache.py,sha256=zwCmCnhr-vhBwHqv9i7xMasdBGVNqAGxZvWsncsJn58,5543
+ai_edge_torch/generative/examples/test_models/toy_model_with_external_kv_cache.py,sha256=jmucKpWY_nHEOAh7G62IxpReNmrKWo4PxfELul_h9xQ,5796
 ai_edge_torch/generative/examples/test_models/toy_model_with_kv_cache.py,sha256=lfYUiem_Pbn3vGgPx84BeI8n7rN3-1fImwCLm8Eo2U8,4853
 ai_edge_torch/generative/examples/tiny_llama/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 ai_edge_torch/generative/examples/tiny_llama/convert_to_tflite.py,sha256=nT7Fh-f5ZdwaK3dPoCvZflpJ4fRHjLdFMjk1_uw3-b8,2559
-ai_edge_torch/generative/examples/tiny_llama/tiny_llama.py,sha256=IFRLPG9wz_aLl_zV_6CETCjSM03ukA6bZqqyDLVACuw,5651
+ai_edge_torch/generative/examples/tiny_llama/tiny_llama.py,sha256=to9IlF-X_uIJvO-roZOW1ZMUhmkYbvFjc-tUVaQr6TE,5848
 ai_edge_torch/generative/fx_passes/__init__.py,sha256=aXvYiaHDvETIrh0Q9DDZA_ZBiazGk80DT6nt7lLtC1o,1172
 ai_edge_torch/generative/fx_passes/remove_sdpa_zero_mask_pass.py,sha256=BCAcc_OcEjvbaXQSbc8vlKeMad7E3gCA4BNsUdWRwBI,1966
 ai_edge_torch/generative/layers/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
@@ -94,6 +104,7 @@ ai_edge_torch/generative/quantize/ai_edge_quantizer_glue/__init__.py,sha256=47DE
 ai_edge_torch/generative/quantize/ai_edge_quantizer_glue/translate_recipe.py,sha256=iTNPrlubmq9ia7C3zHl50J2YEMsc4o33GwL5tr5VkkE,5229
 ai_edge_torch/generative/test/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/generative/test/loader_test.py,sha256=N88CbrLW7Q2x1EyurwdXQ6YjsA-ySQcPxpZH3QOGp-M,3317
+ai_edge_torch/generative/test/test_experimental_ekv.py,sha256=qMR0r7Pr_t2bn-cyeA7Qw_Rl94H1NmFcqM2ua8gpDDw,4230
 ai_edge_torch/generative/test/test_model_conversion.py,sha256=LsPTrLC1I4JW2GowTS3V9Eu257vLHr2Yj5f_qaFUX84,7589
 ai_edge_torch/generative/test/test_quantize.py,sha256=QbF7LC9olJFGXqlAVGciac7xXc4rDtCSr71tTIYuqPk,5230
 ai_edge_torch/generative/utilities/__init__.py,sha256=-_jxnnFnCgnTU4oTm4MnRsvL5lqhomBNdFBbqfmfHPo,720
@@ -114,8 +125,8 @@ ai_edge_torch/quantize/quant_config.py,sha256=eO9Ra160ITjQSyRBEGy6nNIVH3gYacSWDd
 ai_edge_torch/testing/__init__.py,sha256=hHLluseD2R0Hh4W6XZRIXY_dRQeYudjsrKGf6LZz65g,671
 ai_edge_torch/testing/model_coverage/__init__.py,sha256=5P8J6Zk5YYtDvTBucFvB9NGSRI7Gw_24WnrbhXgycEE,765
 ai_edge_torch/testing/model_coverage/model_coverage.py,sha256=kzIulTldq8R9E-lAZsvfSTvLu3FYEX7b9DyYM3qisXM,4485
-ai_edge_torch_nightly-0.2.0.dev20240718.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
-ai_edge_torch_nightly-0.2.0.dev20240718.dist-info/METADATA,sha256=r8YZWPZEhL5gi1oIR9sDZppTSZIuxeHH5isLO4NiSj8,1745
-ai_edge_torch_nightly-0.2.0.dev20240718.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
-ai_edge_torch_nightly-0.2.0.dev20240718.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
-ai_edge_torch_nightly-0.2.0.dev20240718.dist-info/RECORD,,
+ai_edge_torch_nightly-0.2.0.dev20240720.dist-info/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
+ai_edge_torch_nightly-0.2.0.dev20240720.dist-info/METADATA,sha256=xkXzcnmvTzJRRNOJ2c8JnWS1ZCofdlZiKsW5sa5sDyM,1745
+ai_edge_torch_nightly-0.2.0.dev20240720.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
+ai_edge_torch_nightly-0.2.0.dev20240720.dist-info/top_level.txt,sha256=5KXRaF2hwkApYxf7Y8y_tVb9aulGTlbOoNdbx1aKRkE,14
+ai_edge_torch_nightly-0.2.0.dev20240720.dist-info/RECORD,,