tpu-inference 0.12.0.dev20251222__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- tests/__init__.py +13 -0
- tests/core/__init__.py +13 -0
- tests/core/test_core_tpu.py +513 -0
- tests/core/test_disagg_executor.py +60 -0
- tests/core/test_disagg_utils.py +67 -0
- tests/core/test_dp_scheduler.py +724 -0
- tests/core/test_init.py +63 -0
- tests/distributed/__init__.py +13 -0
- tests/distributed/test_distributed_utils.py +120 -0
- tests/distributed/test_tpu_connector.py +478 -0
- tests/e2e/__init__.py +13 -0
- tests/e2e/test_async_scheduler.py +211 -0
- tests/e2e/test_data_parallel.py +393 -0
- tests/e2e/test_local_disagg.py +257 -0
- tests/e2e/test_model_loader.py +268 -0
- tests/e2e/test_multi_modal_inference.py +111 -0
- tests/e2e/test_pipeline_parallel.py +265 -0
- tests/e2e/test_runai_model_streamer_loader.py +104 -0
- tests/e2e/test_sampling_params.py +269 -0
- tests/e2e/test_speculative_decoding.py +291 -0
- tests/e2e/test_structured_decoding.py +46 -0
- tests/executors/__init__.py +13 -0
- tests/executors/test_ray_distributed_executor.py +199 -0
- tests/experimental/__init__.py +13 -0
- tests/experimental/test_llama3_jax_stashed.py +208 -0
- tests/kernels/__init__.py +13 -0
- tests/kernels/collectives/__init__.py +13 -0
- tests/kernels/collectives/all_gather_matmul_kernel_test.py +69 -0
- tests/kernels/fused_moe_v1_test.py +388 -0
- tests/kernels/gmm_test.py +205 -0
- tests/kernels/mla_v1_test.py +498 -0
- tests/kernels/quantized_matmul_kernel_test.py +159 -0
- tests/kernels/ragged_kv_cache_update_v2_test.py +248 -0
- tests/kernels/ragged_paged_attention_kernel_v2_test.py +414 -0
- tests/kernels/ragged_paged_attention_kernel_v3_hd64_test.py +565 -0
- tests/kernels/ragged_paged_attention_kernel_v3_test.py +520 -0
- tests/layers/__init__.py +13 -0
- tests/layers/common/__init__.py +13 -0
- tests/layers/common/test_attention_interface.py +156 -0
- tests/layers/common/test_quantization.py +149 -0
- tests/layers/jax/__init__.py +13 -0
- tests/layers/jax/attention/__init__.py +13 -0
- tests/layers/jax/attention/test_common_attention.py +103 -0
- tests/layers/jax/attention/test_deepseek_v3_attention.py +233 -0
- tests/layers/jax/attention/test_llama4_attention.py +135 -0
- tests/layers/jax/moe/__init__.py +13 -0
- tests/layers/jax/moe/test_deepseek_moe.py +235 -0
- tests/layers/jax/sample/__init__.py +13 -0
- tests/layers/jax/sample/test_rejection_sampler.py +1624 -0
- tests/layers/jax/sample/test_sampling.py +115 -0
- tests/layers/jax/sample/test_sampling_metadata.py +254 -0
- tests/layers/jax/test_layers.py +155 -0
- tests/layers/jax/test_qwix.py +969 -0
- tests/layers/jax/test_rope.py +93 -0
- tests/layers/jax/test_sharding.py +159 -0
- tests/layers/jax/test_transformer_block.py +152 -0
- tests/layers/vllm/__init__.py +13 -0
- tests/layers/vllm/test_attention.py +363 -0
- tests/layers/vllm/test_awq.py +405 -0
- tests/layers/vllm/test_compressed_tensors_moe.py +202 -0
- tests/layers/vllm/test_compressed_tensors_w8a8_fp8.py +403 -0
- tests/layers/vllm/test_compressed_tensors_w8a8_int8.py +426 -0
- tests/layers/vllm/test_fp8.py +17 -0
- tests/layers/vllm/test_mxfp4.py +297 -0
- tests/layers/vllm/test_unquantized.py +621 -0
- tests/layers/vllm/utils.py +72 -0
- tests/lora/__init__.py +13 -0
- tests/lora/conftest.py +46 -0
- tests/lora/test_bgmv.py +57 -0
- tests/lora/test_layers.py +666 -0
- tests/lora/test_lora.py +147 -0
- tests/lora/test_lora_perf.py +67 -0
- tests/lora/utils.py +88 -0
- tests/models/__init__.py +13 -0
- tests/models/common/__init__.py +13 -0
- tests/models/common/test_model_loader.py +455 -0
- tests/models/jax/__init__.py +13 -0
- tests/models/jax/test_deepseek_v3.py +401 -0
- tests/models/jax/test_llama3.py +184 -0
- tests/models/jax/test_llama4.py +298 -0
- tests/models/jax/test_llama_eagle3.py +197 -0
- tests/models/jax/test_llama_guard_4.py +242 -0
- tests/models/jax/test_qwen2.py +172 -0
- tests/models/jax/test_qwen2_5_vl.py +606 -0
- tests/models/jax/test_qwen3.py +169 -0
- tests/models/jax/test_weight_loading.py +180 -0
- tests/models/jax/utils/__init__.py +13 -0
- tests/models/jax/utils/test_multi_modal_utils.py +212 -0
- tests/platforms/__init__.py +13 -0
- tests/platforms/test_tpu_platform.py +54 -0
- tests/runner/__init__.py +13 -0
- tests/runner/test_block_table.py +395 -0
- tests/runner/test_input_batch.py +226 -0
- tests/runner/test_kv_cache.py +220 -0
- tests/runner/test_kv_cache_manager.py +498 -0
- tests/runner/test_multimodal_manager.py +429 -0
- tests/runner/test_persistent_batch_manager.py +84 -0
- tests/runner/test_speculative_decoding_manager.py +368 -0
- tests/runner/test_structured_decoding_manager.py +220 -0
- tests/runner/test_tpu_runner.py +202 -0
- tests/runner/test_tpu_runner_dp.py +1033 -0
- tests/runner/test_tpu_runner_mesh.py +200 -0
- tests/runner/test_utils.py +411 -0
- tests/spec_decode/__init__.py +13 -0
- tests/spec_decode/test_eagle3.py +311 -0
- tests/test_base.py +215 -0
- tests/test_envs.py +280 -0
- tests/test_tpu_info.py +134 -0
- tests/test_utils.py +193 -0
- tests/worker/__init__.py +13 -0
- tests/worker/tpu_worker_test.py +414 -0
- tpu_inference/__init__.py +67 -0
- tpu_inference/core/__init__.py +13 -0
- tpu_inference/core/core_tpu.py +786 -0
- tpu_inference/core/disagg_executor.py +118 -0
- tpu_inference/core/disagg_utils.py +49 -0
- tpu_inference/core/sched/__init__.py +13 -0
- tpu_inference/core/sched/dp_scheduler.py +814 -0
- tpu_inference/distributed/__init__.py +13 -0
- tpu_inference/distributed/jax_parallel_state.py +81 -0
- tpu_inference/distributed/tpu_connector.py +732 -0
- tpu_inference/distributed/utils.py +112 -0
- tpu_inference/env_override.py +9 -0
- tpu_inference/envs.py +191 -0
- tpu_inference/executors/__init__.py +13 -0
- tpu_inference/executors/ray_distributed_executor.py +399 -0
- tpu_inference/experimental/__init__.py +13 -0
- tpu_inference/experimental/llama3_jax_stashed.py +272 -0
- tpu_inference/kernels/__init__.py +13 -0
- tpu_inference/kernels/collectives/__init__.py +13 -0
- tpu_inference/kernels/collectives/all_gather_matmul.py +741 -0
- tpu_inference/kernels/collectives/all_gather_matmul_tuned_block_sizes.py +65 -0
- tpu_inference/kernels/collectives/util.py +47 -0
- tpu_inference/kernels/flash_attention/__init__.py +13 -0
- tpu_inference/kernels/flash_attention/kernel.py +772 -0
- tpu_inference/kernels/fused_moe/__init__.py +13 -0
- tpu_inference/kernels/fused_moe/v1/__init__.py +13 -0
- tpu_inference/kernels/fused_moe/v1/kernel.py +1612 -0
- tpu_inference/kernels/megablox/__init__.py +13 -0
- tpu_inference/kernels/megablox/common.py +54 -0
- tpu_inference/kernels/megablox/gmm.py +646 -0
- tpu_inference/kernels/mla/__init__.py +13 -0
- tpu_inference/kernels/mla/v1/__init__.py +13 -0
- tpu_inference/kernels/mla/v1/kernel.py +1340 -0
- tpu_inference/kernels/quantized_matmul/__init__.py +13 -0
- tpu_inference/kernels/quantized_matmul/kernel.py +456 -0
- tpu_inference/kernels/quantized_matmul/tuned_block_sizes.py +609 -0
- tpu_inference/kernels/quantized_matmul/util.py +58 -0
- tpu_inference/kernels/ragged_paged_attention/__init__.py +13 -0
- tpu_inference/kernels/ragged_paged_attention/v2/__init__.py +13 -0
- tpu_inference/kernels/ragged_paged_attention/v2/kernel.py +876 -0
- tpu_inference/kernels/ragged_paged_attention/v2/ragged_kv_cache_update.py +288 -0
- tpu_inference/kernels/ragged_paged_attention/v2/tuned_block_sizes.py +1482 -0
- tpu_inference/kernels/ragged_paged_attention/v3/__init__.py +13 -0
- tpu_inference/kernels/ragged_paged_attention/v3/kernel.py +1594 -0
- tpu_inference/kernels/ragged_paged_attention/v3/kernel_hd64.py +1586 -0
- tpu_inference/kernels/ragged_paged_attention/v3/tuned_block_sizes.py +4460 -0
- tpu_inference/kernels/ragged_paged_attention/v3/tuned_block_sizes_hd64.py +548 -0
- tpu_inference/kernels/ragged_paged_attention/v3/util.py +65 -0
- tpu_inference/layers/__init__.py +13 -0
- tpu_inference/layers/common/__init__.py +13 -0
- tpu_inference/layers/common/attention_interface.py +403 -0
- tpu_inference/layers/common/attention_metadata.py +48 -0
- tpu_inference/layers/common/binary_search.py +295 -0
- tpu_inference/layers/common/quant_methods.py +23 -0
- tpu_inference/layers/common/quantization.py +270 -0
- tpu_inference/layers/common/sharding.py +600 -0
- tpu_inference/layers/jax/__init__.py +13 -0
- tpu_inference/layers/jax/attention/__init__.py +13 -0
- tpu_inference/layers/jax/attention/attention.py +268 -0
- tpu_inference/layers/jax/attention/deepseek_v3_attention.py +547 -0
- tpu_inference/layers/jax/attention/gpt_oss_attention.py +275 -0
- tpu_inference/layers/jax/attention/llama4_attention.py +167 -0
- tpu_inference/layers/jax/base.py +165 -0
- tpu_inference/layers/jax/constants.py +101 -0
- tpu_inference/layers/jax/layers.py +315 -0
- tpu_inference/layers/jax/misc.py +30 -0
- tpu_inference/layers/jax/moe/__init__.py +13 -0
- tpu_inference/layers/jax/moe/deepseek_v3_moe.py +615 -0
- tpu_inference/layers/jax/moe/gpt_oss_moe.py +199 -0
- tpu_inference/layers/jax/moe/moe.py +249 -0
- tpu_inference/layers/jax/pp_utils.py +53 -0
- tpu_inference/layers/jax/rope.py +294 -0
- tpu_inference/layers/jax/rope_interface.py +228 -0
- tpu_inference/layers/jax/sample/__init__.py +13 -0
- tpu_inference/layers/jax/sample/rejection_sampler.py +528 -0
- tpu_inference/layers/jax/sample/sampling.py +110 -0
- tpu_inference/layers/jax/sample/sampling_metadata.py +90 -0
- tpu_inference/layers/jax/transformer_block.py +121 -0
- tpu_inference/layers/vllm/__init__.py +13 -0
- tpu_inference/layers/vllm/attention.py +221 -0
- tpu_inference/layers/vllm/fused_moe.py +502 -0
- tpu_inference/layers/vllm/linear_common.py +221 -0
- tpu_inference/layers/vllm/quantization/__init__.py +55 -0
- tpu_inference/layers/vllm/quantization/awq.py +221 -0
- tpu_inference/layers/vllm/quantization/common.py +124 -0
- tpu_inference/layers/vllm/quantization/compressed_tensors/__init__.py +13 -0
- tpu_inference/layers/vllm/quantization/compressed_tensors/compressed_tensors.py +135 -0
- tpu_inference/layers/vllm/quantization/compressed_tensors/compressed_tensors_moe.py +266 -0
- tpu_inference/layers/vllm/quantization/compressed_tensors/schemes/__init__.py +13 -0
- tpu_inference/layers/vllm/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_fp8.py +222 -0
- tpu_inference/layers/vllm/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_int8.py +150 -0
- tpu_inference/layers/vllm/quantization/fp8.py +118 -0
- tpu_inference/layers/vllm/quantization/mxfp4.py +396 -0
- tpu_inference/layers/vllm/quantization/unquantized.py +416 -0
- tpu_inference/layers/vllm/sharding.py +244 -0
- tpu_inference/logger.py +10 -0
- tpu_inference/lora/__init__.py +13 -0
- tpu_inference/lora/torch_lora_ops.py +98 -0
- tpu_inference/lora/torch_punica_tpu.py +310 -0
- tpu_inference/models/__init__.py +13 -0
- tpu_inference/models/common/__init__.py +13 -0
- tpu_inference/models/common/model_loader.py +520 -0
- tpu_inference/models/jax/__init__.py +13 -0
- tpu_inference/models/jax/deepseek_v3.py +978 -0
- tpu_inference/models/jax/gpt_oss.py +508 -0
- tpu_inference/models/jax/jax_intermediate_tensor.py +93 -0
- tpu_inference/models/jax/llama3.py +436 -0
- tpu_inference/models/jax/llama4.py +643 -0
- tpu_inference/models/jax/llama_eagle3.py +350 -0
- tpu_inference/models/jax/llama_guard_4.py +375 -0
- tpu_inference/models/jax/qwen2.py +390 -0
- tpu_inference/models/jax/qwen2_5_vl.py +1232 -0
- tpu_inference/models/jax/qwen3.py +318 -0
- tpu_inference/models/jax/utils/__init__.py +13 -0
- tpu_inference/models/jax/utils/file_utils.py +110 -0
- tpu_inference/models/jax/utils/multi_modal_utils.py +177 -0
- tpu_inference/models/jax/utils/qwix/__init__.py +13 -0
- tpu_inference/models/jax/utils/qwix/qwix_utils.py +713 -0
- tpu_inference/models/jax/utils/weight_utils.py +621 -0
- tpu_inference/models/vllm/__init__.py +13 -0
- tpu_inference/models/vllm/vllm_model_wrapper.py +307 -0
- tpu_inference/models/vllm/vllm_model_wrapper_context.py +59 -0
- tpu_inference/platforms/__init__.py +16 -0
- tpu_inference/platforms/tpu_platform.py +258 -0
- tpu_inference/runner/__init__.py +13 -0
- tpu_inference/runner/block_table.py +122 -0
- tpu_inference/runner/compilation_manager.py +890 -0
- tpu_inference/runner/input_batch.py +435 -0
- tpu_inference/runner/kv_cache.py +166 -0
- tpu_inference/runner/kv_cache_manager.py +508 -0
- tpu_inference/runner/lora_utils.py +106 -0
- tpu_inference/runner/multimodal_manager.py +231 -0
- tpu_inference/runner/persistent_batch_manager.py +296 -0
- tpu_inference/runner/speculative_decoding_manager.py +262 -0
- tpu_inference/runner/structured_decoding_manager.py +101 -0
- tpu_inference/runner/tpu_runner.py +1768 -0
- tpu_inference/runner/utils.py +426 -0
- tpu_inference/spec_decode/__init__.py +13 -0
- tpu_inference/spec_decode/jax/__init__.py +13 -0
- tpu_inference/spec_decode/jax/eagle3.py +430 -0
- tpu_inference/tpu_info.py +92 -0
- tpu_inference/utils.py +345 -0
- tpu_inference/worker/__init__.py +13 -0
- tpu_inference/worker/tpu_worker.py +468 -0
- tpu_inference-0.12.0.dev20251222.dist-info/METADATA +106 -0
- tpu_inference-0.12.0.dev20251222.dist-info/RECORD +260 -0
- tpu_inference-0.12.0.dev20251222.dist-info/WHEEL +5 -0
- tpu_inference-0.12.0.dev20251222.dist-info/licenses/LICENSE +201 -0
- tpu_inference-0.12.0.dev20251222.dist-info/top_level.txt +2 -0
|
@@ -0,0 +1,65 @@
|
|
|
1
|
+
# SPDX-License-Identifier: Apache-2.0
|
|
2
|
+
"""All-gather matmul kernel's tuned block sizes."""
|
|
3
|
+
|
|
4
|
+
import re
|
|
5
|
+
|
|
6
|
+
import jax
|
|
7
|
+
|
|
8
|
+
# key:
|
|
9
|
+
# - tpu_version
|
|
10
|
+
# - m
|
|
11
|
+
# - n
|
|
12
|
+
# - k
|
|
13
|
+
# - dtype
|
|
14
|
+
# - tp_size
|
|
15
|
+
# value:
|
|
16
|
+
# - bn
|
|
17
|
+
# - bk
|
|
18
|
+
TUNED_BLOCK_SIZES = {
|
|
19
|
+
# go/keep-sorted start
|
|
20
|
+
(6, 1024, 51200, 5120, 'bfloat16', 8): (6400, 2560),
|
|
21
|
+
(6, 1024, 57344, 8192, 'bfloat16', 8): (7168, 8192),
|
|
22
|
+
(6, 2048, 51200, 5120, 'bfloat16', 8): (1280, 5120),
|
|
23
|
+
(6, 2048, 57344, 8192, 'bfloat16', 8): (1024, 8192),
|
|
24
|
+
(6, 4096, 51200, 5120, 'bfloat16', 8): (3200, 5120),
|
|
25
|
+
(6, 8192, 51200, 5120, 'bfloat16', 8): (1280, 5120),
|
|
26
|
+
# go/keep-sorted end
|
|
27
|
+
}
|
|
28
|
+
|
|
29
|
+
|
|
30
|
+
def get_tpu_version() -> int:
|
|
31
|
+
"""Returns the numeric version of the TPU, or -1 if not on TPU."""
|
|
32
|
+
kind = jax.devices()[0].device_kind
|
|
33
|
+
if 'TPU' not in kind:
|
|
34
|
+
return -1
|
|
35
|
+
if kind.endswith(' lite'):
|
|
36
|
+
kind = kind[:-len(' lite')]
|
|
37
|
+
|
|
38
|
+
# v6: "TPU v6"
|
|
39
|
+
# v7: "TPU7x"
|
|
40
|
+
assert kind[:3] == 'TPU', kind
|
|
41
|
+
return int(re.search(r'\d+', kind).group())
|
|
42
|
+
|
|
43
|
+
|
|
44
|
+
def get_key(
|
|
45
|
+
m,
|
|
46
|
+
n,
|
|
47
|
+
k,
|
|
48
|
+
dtype,
|
|
49
|
+
tp_size,
|
|
50
|
+
):
|
|
51
|
+
"""Returns the key for the given parameters."""
|
|
52
|
+
return (
|
|
53
|
+
get_tpu_version(),
|
|
54
|
+
m,
|
|
55
|
+
n,
|
|
56
|
+
k,
|
|
57
|
+
dtype,
|
|
58
|
+
tp_size,
|
|
59
|
+
)
|
|
60
|
+
|
|
61
|
+
|
|
62
|
+
def get_tuned_block_sizes(m, n, k, dtype_name, tp_size):
|
|
63
|
+
"""Returns the tuned block sizes for the given parameters."""
|
|
64
|
+
key = get_key(m, n, k, dtype_name, tp_size)
|
|
65
|
+
return TUNED_BLOCK_SIZES.get(key, (None, None))
|
|
@@ -0,0 +1,47 @@
|
|
|
1
|
+
# SPDX-License-Identifier: Apache-2.0
|
|
2
|
+
"""utilities for collective kernels."""
|
|
3
|
+
|
|
4
|
+
import functools
|
|
5
|
+
|
|
6
|
+
from jax.experimental import pallas as pl
|
|
7
|
+
from jax.experimental.pallas import tpu as pltpu
|
|
8
|
+
|
|
9
|
+
|
|
10
|
+
def local_barrier(left_neighbor, right_neighbor, double_barrier=True):
|
|
11
|
+
"""Performs a barrier with neighbors on the global barrier semaphore.
|
|
12
|
+
|
|
13
|
+
Optionally performs a second barrier, which prevents a potential race
|
|
14
|
+
when reusing the same collective_id across kernel invocations.
|
|
15
|
+
|
|
16
|
+
Args:
|
|
17
|
+
left_neighbor: Left neighbor device id.
|
|
18
|
+
right_neighbor: Right neighbor device id.
|
|
19
|
+
double_barrier: Whether to perform a second barrier.
|
|
20
|
+
"""
|
|
21
|
+
barrier_sem = pltpu.get_barrier_semaphore()
|
|
22
|
+
for neighbor in [left_neighbor, right_neighbor]:
|
|
23
|
+
pltpu.semaphore_signal(
|
|
24
|
+
barrier_sem,
|
|
25
|
+
inc=1,
|
|
26
|
+
device_id=(neighbor, ),
|
|
27
|
+
device_id_type=pltpu.DeviceIdType.MESH,
|
|
28
|
+
)
|
|
29
|
+
pltpu.semaphore_wait(barrier_sem, 2)
|
|
30
|
+
if double_barrier:
|
|
31
|
+
# The double-barrier prevents a race condition where one neighbor can
|
|
32
|
+
# re-enter the kernel again on a subsequent call and increment the
|
|
33
|
+
# barrier semaphore a second time. This would unblock the current device
|
|
34
|
+
# even if the other neighbor is not ready yet.
|
|
35
|
+
# To implement a double-barrier, we stack-allocate a second REGULAR
|
|
36
|
+
# semaphore using run_scoped.
|
|
37
|
+
@functools.partial(pl.run_scoped,
|
|
38
|
+
second_barrier=pltpu.SemaphoreType.REGULAR)
|
|
39
|
+
def _(second_barrier):
|
|
40
|
+
for neighbor in [left_neighbor, right_neighbor]:
|
|
41
|
+
pltpu.semaphore_signal(
|
|
42
|
+
second_barrier,
|
|
43
|
+
inc=1,
|
|
44
|
+
device_id=(neighbor, ),
|
|
45
|
+
device_id_type=pltpu.DeviceIdType.MESH,
|
|
46
|
+
)
|
|
47
|
+
pltpu.semaphore_wait(second_barrier, 2)
|
|
@@ -0,0 +1,13 @@
|
|
|
1
|
+
# Copyright 2025 Google LLC
|
|
2
|
+
#
|
|
3
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
4
|
+
# you may not use this file except in compliance with the License.
|
|
5
|
+
# You may obtain a copy of the License at
|
|
6
|
+
#
|
|
7
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
|
8
|
+
#
|
|
9
|
+
# Unless required by applicable law or agreed to in writing, software
|
|
10
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
11
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
12
|
+
# See the License for the specific language governing permissions and
|
|
13
|
+
# limitations under the License.
|