tpu-inference 0.12.0.dev20251222__py3-none-any.whl

tests/experimental/test_llama3_jax_stashed.py

@@ -0,0 +1,208 @@
+# 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 types import SimpleNamespace
+from unittest.mock import MagicMock, patch
+
+import jax
+import jax.numpy as jnp
+import numpy as np
+import pytest
+from flax import nnx
+from flax.typing import PRNGKey
+from jax.sharding import Mesh
+
+from tpu_inference.experimental.llama3_jax_stashed import (Llama3WeightLoader,
+                                                           LlamaForCausalLM)
+
+
+class MockParam:
+    """A mock for a parameter used in the Llama model."""
+
+    def __init__(self, shape=(32, 128)):
+        self.value = SimpleNamespace(shape=shape)
+        # The sharding spec is accessed during weight loading
+        self.sharding = SimpleNamespace(spec=None)
+
+    # Allow the mock parameter's value to be updated
+    def __setattr__(self, name, value):
+        if name == "value":
+            self.__dict__[name] = value
+        else:
+            super().__setattr__(name, value)
+
+
+class MockVllmConfig:
+    """A mock VllmConfig sufficient for testing the Llama3 model."""
+
+    def __init__(self,
+                 model_name: str,
+                 random_weights: bool = False,
+                 tensor_parallelism: int = 1):
+        self.model_config = SimpleNamespace(model=model_name,
+                                            dtype="bfloat16",
+                                            hf_overrides={},
+                                            override_generation_config={})
+        self.load_config = MagicMock()
+        self.additional_config = {
+            "random_weights": random_weights,
+            "sharding": {
+                "sharding_strategy": {
+                    "tensor_parallelism": tensor_parallelism
+                }
+            }
+        }
+
+        # NOTE (jacobplatin): we could add a quantized KV cache test, but
+        # we'll skip it for now.
+        self.cache_config = MagicMock(cache_dtype="auto")
+
+
+@pytest.fixture(scope="module")
+def mesh():
+    """
+    Creates a mesh with all required axes for testing.
+    FIX: The sharding logic expects 'data', 'model', and 'expert' axes.
+    This creates a 3D mesh to satisfy the sharding rules, even on a single device.
+    """
+    if not jax.devices():
+        pytest.skip("No JAX devices available for mesh creation.")
+
+    devices = np.array(jax.local_devices())
+    # Reshape devices into a 3D array to name 3 axes: data, model, and expert.
+    # The 'model' and 'expert' axes will have a size of 1.
+    num_devices = len(devices)
+    device_mesh = devices.reshape((num_devices, 1, 1))
+
+    with Mesh(device_mesh, axis_names=('data', 'model', 'expert')) as m:
+        yield m
+
+
+@pytest.fixture
+def rng() -> PRNGKey:
+    """Provides a reusable JAX PRNGKey."""
+    return jax.random.PRNGKey(42)
+
+
+@pytest.fixture
+def mock_vllm_config_8b() -> MockVllmConfig:
+    return MockVllmConfig(model_name="meta-llama/Llama-3-8B")
+
+
+@pytest.fixture
+def mock_vllm_config_70b() -> MockVllmConfig:
+    return MockVllmConfig(model_name="meta-llama/Llama-3-70B-Instruct")
+
+
+@pytest.fixture
+def mock_vllm_config_unknown() -> MockVllmConfig:
+    return MockVllmConfig(model_name="some-other-model")
+
+
+# --- Test Cases ---
+
+
+class TestLlamaForCausalLM:
+    """Tests for the main LlamaForCausalLM model class."""
+
+    def test_init_8b_variant(self, mock_vllm_config_8b, rng, mesh):
+        """Tests correct parameter detection for the 8B model variant."""
+        model = LlamaForCausalLM(mock_vllm_config_8b, rng, mesh)
+        assert model.hidden_size == 4096
+        assert "8b" in model.vllm_config.model_config.model.lower()
+
+    def test_init_70b_variant(self, mock_vllm_config_70b, rng, mesh):
+        """Tests correct parameter detection for the 70B model variant."""
+        model = nnx.eval_shape(
+            lambda: LlamaForCausalLM(mock_vllm_config_70b, rng, mesh))
+        assert model.hidden_size == 8192
+        assert "70b" in model.vllm_config.model_config.model.lower()
+
+    def test_init_unknown_variant_raises_error(self, mock_vllm_config_unknown,
+                                               rng, mesh):
+        """Tests that an unknown model variant raises a ValueError."""
+        with pytest.raises(ValueError,
+                           match="Could not determine Llama3 variant"):
+            LlamaForCausalLM(mock_vllm_config_unknown, rng, mesh)
+
+    def test_create_model_with_random_weights(self, mock_vllm_config_8b, rng,
+                                              mesh):
+        """
+        Tests that random weight initialization creates concrete, non-zero-variance arrays.
+        """
+        with jax.set_mesh(mesh):
+            model = LlamaForCausalLM(vllm_config=mock_vllm_config_8b,
+                                     rng=rng,
+                                     mesh=mesh,
+                                     force_random_weights=True)
+
+            embedding_weight = model.embedder.input_embedding_table_VD.value
+            attention_q_kernel = model.layers[0].attn.kernel_q_proj_DNH.value
+            final_norm_scale = model.final_norm.scale.value
+
+            assert isinstance(embedding_weight, jax.Array)
+            assert isinstance(attention_q_kernel, jax.Array)
+            assert isinstance(final_norm_scale, jax.Array)
+
+            assert jnp.std(embedding_weight) > 0
+            assert jnp.std(attention_q_kernel) > 0
+
+            assert jnp.all(final_norm_scale == 1.0)
+
+    @patch("tpu_inference.experimental.llama3_jax_stashed.Llama3WeightLoader")
+    def test_load_weights_called_correctly(self, mock_loader_cls, rng, mesh):
+        """Tests that the weight loader is called correctly for checkpoint loading."""
+        vllm_config = MockVllmConfig(model_name="llama3-8b",
+                                     random_weights=False)
+        model = LlamaForCausalLM(vllm_config, rng, mesh)
+
+        mock_loader_instance = MagicMock()
+        mock_loader_cls.return_value = mock_loader_instance
+        model.load_weights(rng, cache_dir="/tmp/cache")
+        mock_loader_cls.assert_called_once_with(vllm_config=vllm_config,
+                                                hidden_size=4096,
+                                                attn_heads=32,
+                                                num_key_value_heads=8,
+                                                attn_head_dim=128)
+        mock_loader_instance.load_weights.assert_called_once_with(model)
+
+
+class TestLlama3WeightLoader:
+    """Tests for the Llama3WeightLoader class."""
+
+    @pytest.fixture
+    def weight_loader(self):
+        # Patch the superclass's setup to isolate the Llama3 loader's logic
+        return Llama3WeightLoader(vllm_config=MockVllmConfig("test-model"),
+                                  hidden_size=32,
+                                  attn_heads=4,
+                                  num_key_value_heads=2,
+                                  attn_head_dim=8)
+
+    def test_load_weights_transformation(self, weight_loader, rng, mesh):
+        """Tests that weights are correctly reshaped, transposed, and loaded."""
+        vllm_config = MockVllmConfig("llama3-8b-small-test",
+                                     random_weights=False)
+
+        # Create a model instance but override its config for the test.
+        model = LlamaForCausalLM(vllm_config, rng, mesh)
+
+        with patch(
+                "tpu_inference.experimental.llama3_jax_stashed.load_hf_weights"
+        ) as mock_load:
+            # This will now pass after the code fix
+            weight_loader.load_weights(model)
+
+            # Assert that shard_put was called with the correctly transposed weight
+            mock_load.assert_called_once()

tests/kernels/__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/kernels/collectives/__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/kernels/collectives/all_gather_matmul_kernel_test.py

@@ -0,0 +1,69 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import os
+
+import jax
+import jax.numpy as jnp
+from absl.testing import absltest, parameterized
+from jax._src import test_util as jtu
+
+from tpu_inference import utils
+from tpu_inference.kernels.collectives import all_gather_matmul
+
+jax.config.parse_flags_with_absl()
+
+P = jax.sharding.PartitionSpec
+
+SpongeDir: str | None = os.environ.get('TEST_UNDECLARED_OUTPUTS_DIR', None)
+
+
+@jtu.with_config(jax_numpy_dtype_promotion='standard')
+class AllGatherMatmulTest(jtu.JaxTestCase):
+
+    @parameterized.product(
+        grid_k=[1, 2, 3],
+        grid_n=[1, 2, 3],
+        rhs_transpose=[True, False],
+    )
+    def test_all_gather_matmul(self, grid_k, grid_n, rhs_transpose):
+        if jax.device_count() != 8:
+            self.skipTest('Not enough devices for test')
+
+        axis_name = 'x'
+        num_devices = jax.device_count()
+        mesh = utils.make_optimized_mesh((num_devices, ), (axis_name, ))
+        bk, bn = 1024, 1024
+        m, k, n = 1024, bk * grid_k, bn * grid_n * num_devices
+
+        # Run the test 10 times to expose race conditions as much as possible.
+        for i in range(10):
+            # Create input data
+            prng_key = jax.random.key(1234 + i)
+            k0, k1 = jax.random.split(prng_key, 2)
+            x = jax.random.normal(k0, (m, k), dtype=jnp.bfloat16)
+            y_shape = (n, k) if rhs_transpose else (k, n)
+            y_sharding = P(axis_name, None) if rhs_transpose else P(
+                None, axis_name)
+            y = jax.random.normal(k1, y_shape, dtype=jnp.bfloat16)
+            sharded_x = jax.device_put(
+                x, jax.sharding.NamedSharding(mesh, P(axis_name, None)))
+            sharded_y = jax.device_put(
+                y, jax.sharding.NamedSharding(mesh, y_sharding))
+
+            # Run the all_gather_matmul function
+            output = all_gather_matmul.all_gather_matmul(
+                sharded_x,
+                sharded_y,
+                mesh,
+                axis_name,
+                bk=bk,
+                bn=bn,
+                rhs_transpose=rhs_transpose,
+            )
+            y_for_dot = sharded_y.T if rhs_transpose else sharded_y
+            expected_output = jnp.dot(sharded_x, y_for_dot)
+            self.assertAllClose(output, expected_output, atol=1e-2, rtol=1e-2)
+
+
+if __name__ == "__main__":
+    absltest.main(testLoader=jtu.JaxTestLoader())

tests/kernels/fused_moe_v1_test.py

@@ -0,0 +1,388 @@
+# 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 jax
+import jax.numpy as jnp
+import numpy as np
+from absl.testing import absltest, parameterized
+from jax._src import test_util as jtu
+from jax.sharding import Mesh
+
+from tpu_inference.kernels.fused_moe.v1.kernel import fused_ep_moe, ref_moe
+
+jax.config.parse_flags_with_absl()
+
+
+def cdiv(a, b):
+    assert b != 0
+    return (a + b - 1) // b
+
+
+def align_to(x, a):
+    return cdiv(x, a) * a
+
+
+def gen_moe_inputs(
+    dtype,
+    top_k,
+    num_experts,
+    hidden_size,
+    intermediate_size,
+    num_tokens,
+    *,
+    seed=1234,
+    has_bias=False,
+):
+    key = jax.random.key(seed)
+    k0, k1, k2, k3, k4, k5, k6 = jax.random.split(key, 7)
+
+    a = jax.random.normal(k0, (num_tokens, hidden_size),
+                          dtype=jnp.float32).astype(dtype) / 10
+
+    w1 = (jax.random.normal(
+        k1,
+        (num_experts, 2, hidden_size, intermediate_size),
+        dtype=jnp.float32,
+    ) / 10).astype(dtype)
+    w2 = (jax.random.normal(k2, (num_experts, intermediate_size, hidden_size),
+                            dtype=jnp.float32) / 10).astype(dtype)
+
+    if has_bias:
+        b1 = (jax.random.normal(k3, (num_experts, 2, intermediate_size),
+                                dtype=jnp.float32) / 10).astype(dtype)
+        b2 = (jax.random.normal(k4, (num_experts, hidden_size),
+                                dtype=jnp.float32) / 10).astype(dtype)
+    else:
+        b1 = b2 = None
+
+    gating_output = (
+        jax.random.normal(k5, (num_tokens, num_experts), dtype=jnp.float32) +
+        jnp.arange(num_tokens * num_experts, dtype=jnp.float32).reshape(
+            num_tokens, num_experts) / 100)
+
+    # To generate unique top-k!
+    top_k_indices = jax.random.randint(k6, (num_tokens, top_k),
+                                       minval=0,
+                                       maxval=num_experts - 1,
+                                       dtype=jnp.int32)
+
+    one_hot = (jnp.sum(
+        jax.nn.one_hot(top_k_indices, num_experts, dtype=jnp.float32),
+        axis=1,
+    ) * 30)
+
+    gating_output = (gating_output + one_hot).astype(dtype)
+
+    return a, w1, w2, b1, b2, gating_output
+
+
+def sub_channel_quantize(x, quant_dtype, wsz=256):
+    """Quantizes x with sub-channel quantization on the 2nd minor."""
+    if jnp.issubdtype(quant_dtype, jnp.floating):
+        dtype_info = jnp.finfo(quant_dtype)
+    else:
+        dtype_info = jnp.iinfo(quant_dtype)
+    dtype_max = float(dtype_info.max)
+    w_lst, scale_lst = [], []
+    assert len(x.shape) >= 2
+    assert x.shape[-2] % wsz == 0
+    for i in range(0, x.shape[-2], wsz):
+        y = x[..., i:i + wsz, :]
+        abs_max = jnp.abs(y).max(axis=-2, keepdims=True)
+        scale = (abs_max / dtype_max).astype(jnp.float32)
+        w = (y / scale).astype(quant_dtype)
+        w_lst.append(w)
+        scale_lst.append(scale)
+    return jnp.concat(w_lst, axis=-2), jnp.concat(scale_lst, axis=-2)
+
+
+@jtu.with_config(jax_numpy_dtype_promotion="standard")
+class MoEKernelTest(jtu.JaxTestCase):
+
+    def setUp(self):
+        super().setUp()
+        self.mesh_devices = sorted(
+            jax.devices(),
+            key=lambda x: (
+                x.coords[0],
+                (-1 if x.coords[0] % 2 else 1) * x.coords[1],
+            ),
+        )
+        self.mesh = Mesh(np.array(self.mesh_devices).reshape(1, -1),
+                         axis_names=("data", "model"))
+
+    def _test_moe(
+        self,
+        dtype,
+        top_k,
+        num_experts,
+        hidden_size,
+        intermediate_size,
+        num_tokens,
+        seed,
+        renormalize_topk_logits,
+        bt,
+        bf,
+        bd1,
+        bd2,
+        btc,
+        bfc,
+        bd1c,
+        bd2c,
+        act_fn="silu",
+        w_dtype=None,
+        subc_quant_wsz=None,
+        has_bias=False,
+        atol=2e-1,
+        rtol=2e-1,
+    ):
+        a, w1, w2, b1, b2, gating_output = gen_moe_inputs(
+            dtype,
+            top_k,
+            num_experts,
+            hidden_size,
+            intermediate_size,
+            num_tokens,
+            seed=seed,
+            has_bias=has_bias,
+        )
+        w1_scale = None
+        w2_scale = None
+        if w_dtype is not None:
+            if subc_quant_wsz is None:
+                subc_quant_wsz = 256
+            w1, w1_scale = sub_channel_quantize(w1, w_dtype, subc_quant_wsz)
+            w2, w2_scale = sub_channel_quantize(w2, w_dtype, subc_quant_wsz)
+
+        actual = fused_ep_moe(
+            mesh=self.mesh,
+            tokens=a,
+            w1=w1,
+            w2=w2,
+            gating_output=gating_output,
+            top_k=top_k,
+            renormalize_topk_logits=renormalize_topk_logits,
+            act_fn=act_fn,
+            subc_quant_wsz=subc_quant_wsz,
+            w1_scale=w1_scale,
+            w2_scale=w2_scale,
+            b1=b1,
+            b2=b2,
+            bt=bt,
+            bf=bf,
+            bd1=bd1,
+            bd2=bd2,
+            btc=btc,
+            bfc=bfc,
+            bd1c=bd1c,
+            bd2c=bd2c,
+        )
+        expected = ref_moe(
+            a,
+            w1,
+            w2,
+            gating_output,
+            top_k,
+            b1=b1,
+            b2=b2,
+            renormalize_topk_logits=renormalize_topk_logits,
+            activation=act_fn,
+            subc_quant_wsz=subc_quant_wsz,
+            w1_scale=w1_scale,
+            w2_scale=w2_scale,
+        )
+        self.assertAllClose(actual, expected, atol=atol, rtol=rtol)
+
+    @parameterized.product(renormalize_topk_logits=[True, False], )
+    def test_basic(self, renormalize_topk_logits):
+        dtype = jnp.bfloat16
+        top_k = 8
+        num_experts = 128
+        hidden_size = 1024
+        intermediate_size = 1024
+        num_tokens = 8 * 32
+        self._test_moe(
+            dtype=dtype,
+            top_k=top_k,
+            num_experts=num_experts,
+            hidden_size=hidden_size,
+            intermediate_size=intermediate_size,
+            num_tokens=num_tokens,
+            seed=1234,
+            renormalize_topk_logits=renormalize_topk_logits,
+            bt=32,
+            bf=1024,
+            bd1=1024,
+            bd2=1024,
+            btc=32,
+            bfc=256,
+            bd1c=256,
+            bd2c=256,
+        )
+
+    @parameterized.product(act_fn=["silu", "gelu", "swigluoai"], )
+    def test_activation(self, act_fn):
+        dtype = jnp.bfloat16
+        top_k = 8
+        num_experts = 128
+        hidden_size = 1024
+        intermediate_size = 1024
+        num_tokens = 8 * 32
+        self._test_moe(
+            dtype=dtype,
+            top_k=top_k,
+            num_experts=num_experts,
+            hidden_size=hidden_size,
+            intermediate_size=intermediate_size,
+            num_tokens=num_tokens,
+            seed=1234,
+            renormalize_topk_logits=True,
+            act_fn=act_fn,
+            bt=32,
+            bf=512,
+            bd1=512,
+            bd2=512,
+            btc=32,
+            bfc=256,
+            bd1c=256,
+            bd2c=256,
+        )
+
+    def test_benchmark_qwen_235(self):
+        num_experts = 128
+        top_k = 8
+        hidden_size = 4096
+        intermediate_size = 1536
+        dtype = jnp.bfloat16
+        num_tokens = 8 * 64
+        seed = 54321
+        renormalize_topk_logits = True
+        self._test_moe(
+            dtype=dtype,
+            top_k=top_k,
+            num_experts=num_experts,
+            hidden_size=hidden_size,
+            intermediate_size=intermediate_size,
+            num_tokens=num_tokens,
+            seed=seed,
+            renormalize_topk_logits=renormalize_topk_logits,
+            bt=64,
+            bf=768,
+            bd1=2048,
+            bd2=2048,
+            btc=64,
+            bfc=768,
+            bd1c=2048,
+            bd2c=2048,
+            act_fn="silu",
+            atol=5e-2,
+            rtol=5e-2,
+        )
+
+    def test_benchmark_qwen_30b_a3b(self):
+        num_experts = 128
+        top_k = 8
+        hidden_size = 2048
+        intermediate_size = 768
+        dtype = jnp.bfloat16
+        num_tokens = 512
+        seed = 54321
+        renormalize_topk_logits = True
+        self._test_moe(
+            dtype=dtype,
+            top_k=top_k,
+            num_experts=num_experts,
+            hidden_size=hidden_size,
+            intermediate_size=intermediate_size,
+            num_tokens=num_tokens,
+            seed=seed,
+            renormalize_topk_logits=renormalize_topk_logits,
+            bt=16,
+            bf=384,
+            bd1=512,
+            bd2=512,
+            btc=16,
+            bfc=384,
+            bd1c=256,
+            bd2c=256,
+            act_fn="silu",
+            atol=5e-2,
+            rtol=5e-2,
+        )
+
+    @parameterized.product(
+        w_dtype=[jnp.int8, jnp.float8_e5m2, jnp.float4_e2m1fn], )
+    def test_sub_channel_quantization(self, w_dtype):
+        if w_dtype in (
+                jnp.float8_e5m2,
+                jnp.float4_e2m1fn,
+        ) and not jtu.is_device_tpu_at_least(version=7):
+            self.skipTest("Expect TPUv7+")
+        dtype = jnp.bfloat16
+        top_k = 8
+        num_experts = 128
+        hidden_size = 1024
+        intermediate_size = 1024
+        num_tokens = 8 * 32
+        self._test_moe(
+            dtype=dtype,
+            top_k=top_k,
+            num_experts=num_experts,
+            hidden_size=hidden_size,
+            intermediate_size=intermediate_size,
+            num_tokens=num_tokens,
+            seed=1234,
+            renormalize_topk_logits=False,
+            w_dtype=w_dtype,
+            subc_quant_wsz=256,
+            bt=32,
+            bf=1024,
+            bd1=1024,
+            bd2=1024,
+            btc=32,
+            bfc=256,
+            bd1c=256,
+            bd2c=256,
+        )
+
+    def test_bias(self):
+        dtype = jnp.bfloat16
+        top_k = 8
+        num_experts = 128
+        hidden_size = 1024
+        intermediate_size = 1024
+        num_tokens = 8 * 32
+        self._test_moe(
+            dtype=dtype,
+            top_k=top_k,
+            num_experts=num_experts,
+            hidden_size=hidden_size,
+            intermediate_size=intermediate_size,
+            num_tokens=num_tokens,
+            seed=1234,
+            renormalize_topk_logits=False,
+            has_bias=True,
+            bt=32,
+            bf=512,
+            bd1=512,
+            bd2=512,
+            btc=32,
+            bfc=256,
+            bd1c=256,
+            bd2c=256,
+        )
+
+
+if __name__ == "__main__":
+    absltest.main(testLoader=jtu.JaxTestLoader())