tpu-inference 0.11.1.dev202512030818__py3-none-any.whl → 0.13.2rc3__py3-none-any.whl

tests/layers/jax/test_transformer_block.py

@@ -0,0 +1,152 @@
+# Copyright 2025 Google LLC
+#
+# 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 unittest
+from unittest.mock import MagicMock
+
+import jax.numpy as jnp
+from flax import nnx
+
+from tpu_inference.layers.jax.layers import DenseFFW
+from tpu_inference.layers.jax.moe.moe import MoE
+from tpu_inference.layers.jax.transformer_block import (
+    SharedExpertsTransformerBlock, TransformerBlock)
+
+
+class TestTransformerBlock(unittest.TestCase):
+    """Unit test suite for the JAX TransformerBlock module."""
+
+    def test_transformer_block_dense_logic(self):
+        """
+        Tests the forward pass logic of a dense TransformerBlock by mocking its sub-modules.
+        This test verifies the sequence of operations and residual connections.
+        """
+        hidden_size = 1024
+
+        mock_pre_attn_norm = MagicMock(spec=nnx.Module)
+        mock_pre_mlp_norm = MagicMock(spec=nnx.Module)
+
+        mock_attn = MagicMock(spec=nnx.Module)
+        dummy_attn_output = jnp.full((64, hidden_size),
+                                     2.0,
+                                     dtype=jnp.bfloat16)
+        dummy_kv_cache = jnp.zeros((8, 16, 16, 128), dtype=jnp.bfloat16)
+        mock_attn.return_value = (dummy_kv_cache, dummy_attn_output)
+
+        mock_mlp = MagicMock(spec=DenseFFW)
+        dummy_mlp_output = jnp.full((64, hidden_size), 3.0, dtype=jnp.bfloat16)
+        mock_mlp.return_value = dummy_mlp_output
+
+        transformer_block = TransformerBlock(
+            pre_attention_norm=mock_pre_attn_norm,
+            pre_mlp_norm=mock_pre_mlp_norm,
+            custom_module=mock_mlp,
+            attn=mock_attn,
+        )
+
+        seq_len = 64
+        x = jnp.ones((seq_len, hidden_size), dtype=jnp.bfloat16)
+        initial_kv_cache = MagicMock()
+        attention_metadata = MagicMock()
+
+        mock_pre_attn_norm.side_effect = lambda val: val
+        mock_pre_mlp_norm.side_effect = lambda val: val
+
+        new_kv_cache, final_output = transformer_block(
+            x,
+            is_prefill=True,
+            kv_cache=initial_kv_cache,
+            attention_metadata=attention_metadata,
+        )
+
+        mock_pre_attn_norm.assert_called_once()
+        self.assertTrue(
+            jnp.array_equal(mock_pre_attn_norm.call_args.args[0], x))
+
+        mock_attn.assert_called_once_with(x, True, initial_kv_cache,
+                                          attention_metadata, True)
+
+        expected_mlp_norm_input = dummy_attn_output + x
+
+        mock_pre_mlp_norm.assert_called_once()
+        self.assertTrue(
+            jnp.array_equal(mock_pre_mlp_norm.call_args.args[0],
+                            expected_mlp_norm_input))
+
+        mock_mlp.assert_called_once()
+        self.assertTrue(
+            jnp.array_equal(mock_mlp.call_args.args[0],
+                            expected_mlp_norm_input))
+
+        expected_final_output = dummy_mlp_output + expected_mlp_norm_input
+        self.assertTrue(jnp.allclose(final_output, expected_final_output))
+
+        self.assertTrue(jnp.array_equal(new_kv_cache, dummy_kv_cache))
+
+    def test_shared_experts_transformer_block_logic(self):
+        """Tests the forward pass logic of a SharedExpertsTransformerBlock."""
+        hidden_size = 1024
+
+        mock_pre_attn_norm = MagicMock(spec=nnx.Module)
+        mock_pre_mlp_norm = MagicMock(spec=nnx.Module)
+
+        mock_attn = MagicMock(spec=nnx.Module)
+        dummy_attn_output = jnp.full((64, hidden_size),
+                                     2.0,
+                                     dtype=jnp.bfloat16)
+        dummy_kv_cache = jnp.zeros((8, 16, 16, 128), dtype=jnp.bfloat16)
+        mock_attn.return_value = (dummy_kv_cache, dummy_attn_output)
+
+        mock_moe = MagicMock(spec=MoE)
+        dummy_moe_output = jnp.full((64, hidden_size), 3.0, dtype=jnp.bfloat16)
+        mock_moe.return_value = dummy_moe_output
+
+        mock_shared_experts = MagicMock(spec=DenseFFW)
+        dummy_shared_experts_output = jnp.full((64, hidden_size),
+                                               4.0,
+                                               dtype=jnp.bfloat16)
+        mock_shared_experts.return_value = dummy_shared_experts_output
+
+        transformer_block = SharedExpertsTransformerBlock(
+            pre_attention_norm=mock_pre_attn_norm,
+            pre_mlp_norm=mock_pre_mlp_norm,
+            custom_module=mock_moe,
+            attn=mock_attn,
+            shared_experts=mock_shared_experts,
+        )
+
+        seq_len = 64
+        x = jnp.ones((seq_len, hidden_size), dtype=jnp.bfloat16)
+        initial_kv_cache = MagicMock()
+        attention_metadata = MagicMock()
+
+        mock_pre_attn_norm.side_effect = lambda val: val
+        mock_pre_mlp_norm.side_effect = lambda val: val
+
+        new_kv_cache, final_output = transformer_block(
+            x,
+            is_prefill=True,
+            kv_cache=initial_kv_cache,
+            attention_metadata=attention_metadata,
+        )
+        self.assertTrue(jnp.array_equal(new_kv_cache, dummy_kv_cache))
+        self.assertEqual(final_output.shape, (seq_len, hidden_size))
+
+        self.assertEqual(mock_moe.call_count, 1)
+        self.assertEqual(mock_attn.call_count, 1)
+        self.assertEqual(mock_shared_experts.call_count, 1)
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/layers/vllm/__init__.py

@@ -0,0 +1,13 @@
+# Copyright 2025 Google LLC
+#
+# 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.

tests/layers/vllm/test_attention.py

@@ -0,0 +1,363 @@
+# Copyright 2025 Google LLC
+#
+# 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.
+
+from unittest.mock import MagicMock
+
+import jax
+import jax.numpy as jnp
+import numpy as np
+import pytest
+import torch
+import torchax
+from jax.sharding import Mesh
+from torchax.interop import torch_view
+from vllm.attention.backends.abstract import AttentionType
+
+from tpu_inference.layers.common.attention_metadata import AttentionMetadata
+from tpu_inference.layers.vllm.attention import (PallasAttentionBackend,
+                                                 PallasAttentionBackendImpl)
+from tpu_inference.models.vllm.vllm_model_wrapper_context import \
+    set_vllm_model_wrapper_context
+from tpu_inference.runner.kv_cache import get_kv_cache_shape_with_mesh
+
+# ---- Test Configuration & Constants ----
+
+# Total number of tokens across all sequences in the batch
+TOTAL_TOKENS = 10
+# Number of sequences in the batch
+NUM_SEQS = 2
+# Padded maximum number of sequences
+MAX_NUM_SEQS = 4
+# Number of attention heads (Query)
+NUM_HEADS = 8
+# Number of attention heads (Key/Value) - for Grouped-Query Attention
+NUM_KV_HEADS = 4
+# Dimension of each attention head
+HEAD_DIM = 128
+# Total number of blocks in the KV cache
+NUM_BLOCKS = 32
+# Number of tokens per block
+BLOCK_SIZE = 16
+# Maximum number of blocks a single sequence can occupy
+MAX_BLOCKS_PER_SEQ = 8
+
+
+def create_inputs(
+    mesh: Mesh,
+    q_dtype: jnp.dtype = jnp.bfloat16,
+    kv_dtype: jnp.dtype = jnp.bfloat16,
+    total_tokens: int = TOTAL_TOKENS,
+    num_seqs: int = NUM_SEQS,
+    max_num_seqs: int = MAX_NUM_SEQS,
+    num_heads: int = NUM_HEADS,
+    num_kv_heads: int = NUM_KV_HEADS,
+    head_dim: int = HEAD_DIM,
+    num_blocks: int = NUM_BLOCKS,
+    block_size: int = BLOCK_SIZE,
+    max_blocks_per_seq: int = MAX_BLOCKS_PER_SEQ,
+):
+    key = jax.random.key(0)
+    q = jax.random.uniform(key, (total_tokens, num_heads * head_dim),
+                           dtype=q_dtype)
+    k = jax.random.uniform(key, (total_tokens, num_kv_heads * head_dim),
+                           dtype=q_dtype)
+    v = jax.random.uniform(key, (total_tokens, num_kv_heads * head_dim),
+                           dtype=q_dtype)
+    q = torch_view(q)
+    k = torch_view(k)
+    v = torch_view(v)
+
+    kv_cache_shape = get_kv_cache_shape_with_mesh(mesh, num_blocks, block_size,
+                                                  num_kv_heads, head_dim,
+                                                  kv_dtype)
+    kv_cache = jax.random.normal(key, kv_cache_shape, dtype=kv_dtype)
+
+    positions = jnp.ones((total_tokens, ), dtype=jnp.int32)
+    block_tables = jnp.zeros((max_num_seqs * max_blocks_per_seq),
+                             dtype=jnp.int32).reshape(-1)
+    seq_lens = jnp.array([5, 5, 0, 0], dtype=jnp.int32)
+    query_start_loc = jnp.array([0, 5, 10, 10, 10], dtype=jnp.int32)
+    request_distribution = jnp.array([0, 0, num_seqs], dtype=jnp.int32)
+
+    metadata = AttentionMetadata(
+        input_positions=positions,
+        block_tables=block_tables,
+        seq_lens=seq_lens,
+        query_start_loc=query_start_loc,
+        request_distribution=request_distribution,
+    )
+
+    return q, k, v, kv_cache, metadata
+
+
+@pytest.fixture
+def mesh():
+    """Provides a mock 1D JAX mesh for testing."""
+    # Create a mesh with available devices, useful for running on CPU/GPU/TPU
+    # For this test, it will likely be a single CPU device.
+    devices = np.array(jax.local_devices())[0:1]
+    if not devices.any():
+        # Add a mock device if no devices are present (e.g., in a CI environment)
+        devices = np.array([jax.devices("cpu")[0]])
+    return Mesh(devices.reshape((-1, 1, 1)), ("data", "attn_dp", "model"))
+
+
+class TestPallasAttentionBackend:
+
+    def test_get_name(self):
+        assert PallasAttentionBackend.get_name() == "PALLAS"
+
+    def test_get_impl_cls(self):
+        assert PallasAttentionBackend.get_impl_cls(
+        ) == PallasAttentionBackendImpl
+
+
+class TestPallasAttentionBackendImpl:
+
+    def test_init_valid_params(self):
+        impl = PallasAttentionBackendImpl(
+            num_heads=32,
+            head_size=128,
+            scale=0.088,
+            num_kv_heads=8,
+            alibi_slopes=None,
+            sliding_window=None,
+            kv_cache_dtype="auto",
+            attn_type=AttentionType.DECODER,
+        )
+
+        assert impl.num_heads == 32
+        assert impl.head_size == 128
+        assert impl.scale == 0.088
+        assert impl.num_kv_heads == 8
+        assert impl.num_queries_per_kv == 4
+        assert impl.sliding_window is None
+
+    def test_init_with_alibi_slopes_raises_error(self):
+        with pytest.raises(NotImplementedError,
+                           match="Alibi slopes is not supported"):
+            PallasAttentionBackendImpl(
+                num_heads=32,
+                head_size=128,
+                scale=0.088,
+                num_kv_heads=8,
+                alibi_slopes=[1.0, 2.0],
+                sliding_window=None,
+                kv_cache_dtype="auto",
+                attn_type=AttentionType.DECODER,
+            )
+
+    def test_init_with_encoder_attention_raises_error(self):
+        with pytest.raises(NotImplementedError,
+                           match="Encoder self-attention"):
+            PallasAttentionBackendImpl(
+                num_heads=32,
+                head_size=128,
+                scale=0.088,
+                num_kv_heads=8,
+                alibi_slopes=None,
+                sliding_window=None,
+                kv_cache_dtype="auto",
+                attn_type=AttentionType.ENCODER,
+            )
+
+    def test_forward(self, mesh):
+        impl = PallasAttentionBackendImpl(
+            num_heads=NUM_HEADS,
+            head_size=HEAD_DIM,
+            scale=0.088,
+            num_kv_heads=NUM_KV_HEADS,
+            alibi_slopes=None,
+            sliding_window=None,
+            kv_cache_dtype="auto",
+            attn_type=AttentionType.DECODER,
+        )
+
+        layer = MagicMock()
+        layer.layer_name = "0"
+
+        query, key, value, kv_cache, metadata = create_inputs(mesh)
+
+        with torchax.default_env(), set_vllm_model_wrapper_context(
+                kv_caches=[kv_cache],
+                mesh=mesh,
+                layer_name_to_kvcache_index={'0': 0}):
+            impl.forward(layer, query, key, value, torch.tensor([]), metadata)
+
+    def test_forward_with_fp8_kv_cache(self, mesh):
+        impl = PallasAttentionBackendImpl(
+            num_heads=NUM_HEADS,
+            head_size=HEAD_DIM,
+            scale=0.088,
+            num_kv_heads=NUM_KV_HEADS,
+            alibi_slopes=None,
+            sliding_window=None,
+            kv_cache_dtype="fp8",
+            attn_type=AttentionType.DECODER,
+        )
+
+        layer = MagicMock()
+        layer.layer_name = "0"
+        layer._q_scale_float = None
+        layer._k_scale_float = 1
+        layer._v_scale_float = 1
+
+        query, key, value, kv_cache, metadata = create_inputs(
+            mesh, kv_dtype=jnp.float8_e4m3fn)
+
+        with torchax.default_env(), set_vllm_model_wrapper_context(
+                kv_caches=[kv_cache],
+                mesh=mesh,
+                layer_name_to_kvcache_index={'0': 0}):
+            impl.forward(layer, query, key, value, torch.tensor([]), metadata)
+
+    def test_forward_with_w8a8(self, mesh):
+        impl = PallasAttentionBackendImpl(
+            num_heads=NUM_HEADS,
+            head_size=HEAD_DIM,
+            scale=0.088,
+            num_kv_heads=NUM_KV_HEADS,
+            alibi_slopes=None,
+            sliding_window=None,
+            kv_cache_dtype="fp8",
+            attn_type=AttentionType.DECODER,
+        )
+
+        layer = MagicMock()
+        layer.layer_name = "0"
+        layer._q_scale_float = 1
+        layer._k_scale_float = 1
+        layer._v_scale_float = 1
+
+        query, key, value, kv_cache, metadata = create_inputs(
+            mesh, kv_dtype=jnp.float8_e4m3fn)
+
+        with torchax.default_env(), set_vllm_model_wrapper_context(
+                kv_caches=[kv_cache],
+                mesh=mesh,
+                layer_name_to_kvcache_index={'0': 0}):
+            impl.forward(layer, query, key, value, torch.tensor([]), metadata)
+
+    def test_forward_with_vllm_kv_cache_raises_error(self, mesh):
+        impl = PallasAttentionBackendImpl(
+            num_heads=NUM_HEADS,
+            head_size=HEAD_DIM,
+            scale=0.088,
+            num_kv_heads=NUM_KV_HEADS,
+            alibi_slopes=None,
+            sliding_window=None,
+            kv_cache_dtype="auto",
+            attn_type=AttentionType.DECODER,
+        )
+
+        layer = MagicMock()
+        layer.layer_name = "0"
+
+        query, key, value, kv_cache, metadata = create_inputs(mesh)
+
+        with torchax.default_env(), set_vllm_model_wrapper_context(
+                kv_caches=[kv_cache],
+                mesh=mesh), pytest.raises(RuntimeError,
+                                          match="should be empty but has"):
+            impl.forward(layer, query, key, value, torch.tensor([1]), metadata)
+
+    def test_forward_with_output_scale_raises_error(self, mesh):
+        impl = PallasAttentionBackendImpl(
+            num_heads=NUM_HEADS,
+            head_size=HEAD_DIM,
+            scale=0.088,
+            num_kv_heads=NUM_KV_HEADS,
+            alibi_slopes=None,
+            sliding_window=None,
+            kv_cache_dtype="auto",
+            attn_type=AttentionType.DECODER,
+        )
+
+        layer = MagicMock()
+        layer.layer_name = "0"
+
+        query, key, value, kv_cache, metadata = create_inputs(mesh)
+        output_scale = torch.tensor([1.0])
+
+        with torchax.default_env(), set_vllm_model_wrapper_context(
+                kv_caches=[kv_cache],
+                mesh=mesh), pytest.raises(NotImplementedError,
+                                          match="fused output quantization"):
+            impl.forward(layer,
+                         query,
+                         key,
+                         value,
+                         torch.tensor([]),
+                         metadata,
+                         output_scale=output_scale)
+
+    def test_forward_with_attention_sink(self, mesh):
+        head_dim = 64
+        sinks = torch.rand([NUM_HEADS], dtype=torch.float32)
+
+        impl = PallasAttentionBackendImpl(num_heads=NUM_HEADS,
+                                          head_size=head_dim,
+                                          scale=0.088,
+                                          num_kv_heads=NUM_KV_HEADS,
+                                          alibi_slopes=None,
+                                          sliding_window=None,
+                                          kv_cache_dtype="auto",
+                                          attn_type=AttentionType.DECODER,
+                                          sinks=sinks)
+        impl.process_weights_after_loading(torch.bfloat16)
+
+        layer = MagicMock()
+        layer.layer_name = "0"
+
+        query, key, value, kv_cache, metadata = create_inputs(
+            mesh, head_dim=head_dim)
+
+        with torchax.default_env(), set_vllm_model_wrapper_context(
+                kv_caches=[kv_cache],
+                mesh=mesh,
+                layer_name_to_kvcache_index={'0': 0}):
+            assert impl.sinks is not None
+            impl.forward(layer, query, key, value, torch.tensor([]), metadata)
+
+    def test_forward_with_attention_sink_head_dim_128_raises_error(self, mesh):
+        head_dim = 128
+        sinks = torch.rand([NUM_HEADS], dtype=torch.float32)
+
+        impl = PallasAttentionBackendImpl(num_heads=NUM_HEADS,
+                                          head_size=head_dim,
+                                          scale=0.088,
+                                          num_kv_heads=NUM_KV_HEADS,
+                                          alibi_slopes=None,
+                                          sliding_window=None,
+                                          kv_cache_dtype="auto",
+                                          attn_type=AttentionType.DECODER,
+                                          sinks=sinks)
+        impl.process_weights_after_loading(torch.bfloat16)
+
+        layer = MagicMock()
+        layer.layer_name = "0"
+
+        query, key, value, kv_cache, metadata = create_inputs(
+            mesh, head_dim=head_dim)
+
+        with torchax.default_env(), set_vllm_model_wrapper_context(
+                kv_caches=[kv_cache],
+                mesh=mesh,
+                layer_name_to_kvcache_index={'0': 0}
+        ), pytest.raises(
+                NotImplementedError,
+                match=
+                "Attention sink support is only available when head_dim==64"):
+            assert impl.sinks is not None
+            impl.forward(layer, query, key, value, torch.tensor([]), metadata)