tpu-inference 0.11.1.dev202512030818__py3-none-any.whl → 0.13.0rc2.post7__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.
Potentially problematic release.
This version of tpu-inference might be problematic. Click here for more details.
- tests/__init__.py +13 -0
- tests/core/__init__.py +13 -0
- tests/core/test_disagg_utils.py +14 -0
- tests/core/test_dp_scheduler.py +650 -768
- tests/core/test_init.py +14 -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 +289 -0
- tests/e2e/test_hybrid_kvcache.py +219 -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 +311 -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 +14 -0
- tests/kernels/gmm_test.py +205 -0
- tests/kernels/mla_v1_test.py +143 -41
- tests/kernels/quantized_matmul_kernel_test.py +2 -34
- tests/kernels/ragged_kv_cache_update_v2_test.py +14 -0
- tests/kernels/ragged_paged_attention_kernel_v2_test.py +14 -0
- tests/kernels/ragged_paged_attention_kernel_v3_hd64_test.py +17 -1
- tests/kernels/ragged_paged_attention_kernel_v3_test.py +17 -1
- 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/{test_quantization.py → layers/jax/test_qwix.py} +183 -50
- 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 +418 -0
- tests/layers/vllm/test_compressed_tensors_w8a8_int8.py +441 -0
- tests/layers/vllm/test_fp8.py +17 -0
- tests/layers/vllm/test_mxfp4.py +312 -0
- tests/layers/vllm/test_unquantized.py +651 -0
- tests/layers/vllm/utils.py +87 -0
- tests/lora/__init__.py +13 -0
- tests/lora/conftest.py +14 -0
- tests/lora/test_bgmv.py +14 -0
- tests/lora/test_layers.py +21 -3
- tests/lora/test_lora.py +15 -1
- tests/lora/test_lora_perf.py +67 -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 +605 -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 +261 -0
- tests/runner/test_tpu_runner_dp.py +1099 -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 +14 -0
- tests/test_envs.py +78 -1
- tests/test_tpu_info.py +14 -0
- tests/test_utils.py +1 -43
- tests/worker/__init__.py +13 -0
- tests/worker/tpu_worker_test.py +414 -0
- tpu_inference/__init__.py +14 -0
- tpu_inference/core/__init__.py +13 -0
- tpu_inference/core/sched/__init__.py +13 -0
- tpu_inference/core/sched/dp_scheduler.py +372 -56
- tpu_inference/distributed/__init__.py +13 -0
- tpu_inference/distributed/jax_parallel_state.py +14 -0
- tpu_inference/distributed/tpu_connector.py +14 -9
- tpu_inference/distributed/utils.py +56 -4
- tpu_inference/envs.py +38 -7
- tpu_inference/executors/__init__.py +13 -0
- tpu_inference/executors/ray_distributed_executor.py +17 -0
- tpu_inference/experimental/__init__.py +13 -0
- tpu_inference/experimental/llama3_jax_stashed.py +14 -0
- tpu_inference/kernels/__init__.py +13 -0
- tpu_inference/kernels/collectives/__init__.py +13 -0
- tpu_inference/kernels/collectives/all_gather_matmul.py +12 -6
- tpu_inference/kernels/collectives/all_gather_matmul_tuned_block_sizes.py +7 -2
- tpu_inference/kernels/flash_attention/__init__.py +13 -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 +370 -324
- 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 +117 -145
- tpu_inference/kernels/quantized_matmul/__init__.py +13 -0
- tpu_inference/kernels/quantized_matmul/kernel.py +69 -8
- 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 +2 -1
- tpu_inference/kernels/ragged_paged_attention/v2/ragged_kv_cache_update.py +2 -1
- tpu_inference/kernels/ragged_paged_attention/v3/__init__.py +13 -0
- tpu_inference/kernels/ragged_paged_attention/v3/kernel.py +194 -101
- tpu_inference/kernels/ragged_paged_attention/v3/kernel_hd64.py +95 -78
- tpu_inference/kernels/ragged_paged_attention/v3/tuned_block_sizes.py +3817 -3504
- tpu_inference/kernels/ragged_paged_attention/v3/tuned_block_sizes_hd64.py +376 -195
- tpu_inference/kernels/ragged_paged_attention/v3/util.py +15 -1
- tpu_inference/layers/__init__.py +13 -0
- tpu_inference/layers/common/__init__.py +13 -0
- tpu_inference/layers/common/attention_interface.py +26 -19
- tpu_inference/layers/common/attention_metadata.py +14 -0
- tpu_inference/layers/common/quant_methods.py +15 -0
- tpu_inference/layers/common/quantization.py +270 -0
- tpu_inference/layers/common/sharding.py +28 -5
- tpu_inference/layers/jax/__init__.py +13 -0
- tpu_inference/layers/jax/attention/__init__.py +13 -0
- tpu_inference/layers/jax/attention/attention.py +19 -6
- tpu_inference/layers/jax/attention/deepseek_v3_attention.py +270 -77
- tpu_inference/layers/jax/attention/gpt_oss_attention.py +24 -11
- tpu_inference/layers/jax/attention/llama4_attention.py +17 -4
- tpu_inference/layers/jax/base.py +14 -0
- tpu_inference/layers/jax/constants.py +13 -0
- tpu_inference/layers/jax/layers.py +14 -0
- tpu_inference/layers/jax/misc.py +14 -0
- tpu_inference/layers/jax/moe/__init__.py +13 -0
- tpu_inference/layers/jax/moe/deepseek_v3_moe.py +20 -13
- tpu_inference/layers/jax/moe/gpt_oss_moe.py +14 -0
- tpu_inference/layers/jax/moe/moe.py +43 -3
- tpu_inference/layers/jax/pp_utils.py +53 -0
- tpu_inference/layers/jax/rope.py +14 -0
- tpu_inference/layers/jax/rope_interface.py +14 -0
- tpu_inference/layers/jax/sample/__init__.py +13 -0
- tpu_inference/layers/jax/sample/rejection_sampler.py +13 -0
- tpu_inference/layers/jax/sample/sampling.py +15 -1
- tpu_inference/layers/jax/sample/sampling_metadata.py +14 -0
- tpu_inference/layers/jax/transformer_block.py +14 -0
- tpu_inference/layers/vllm/__init__.py +13 -0
- tpu_inference/layers/vllm/attention.py +4 -4
- tpu_inference/layers/vllm/fused_moe.py +210 -260
- tpu_inference/layers/vllm/linear_common.py +57 -22
- tpu_inference/layers/vllm/quantization/__init__.py +16 -0
- tpu_inference/layers/vllm/quantization/awq.py +15 -1
- tpu_inference/layers/vllm/quantization/common.py +33 -18
- tpu_inference/layers/vllm/quantization/compressed_tensors/__init__.py +13 -0
- tpu_inference/layers/vllm/quantization/compressed_tensors/compressed_tensors.py +18 -3
- tpu_inference/layers/vllm/quantization/compressed_tensors/compressed_tensors_moe.py +211 -148
- 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 +14 -0
- tpu_inference/layers/vllm/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_int8.py +14 -0
- tpu_inference/layers/vllm/quantization/fp8.py +118 -0
- tpu_inference/layers/vllm/quantization/mxfp4.py +278 -209
- tpu_inference/layers/vllm/quantization/unquantized.py +134 -86
- tpu_inference/layers/vllm/sharding.py +21 -4
- tpu_inference/lora/__init__.py +13 -0
- tpu_inference/lora/torch_lora_ops.py +8 -13
- tpu_inference/models/__init__.py +13 -0
- tpu_inference/models/common/__init__.py +13 -0
- tpu_inference/models/common/model_loader.py +74 -35
- tpu_inference/models/jax/__init__.py +13 -0
- tpu_inference/models/jax/deepseek_v3.py +267 -157
- tpu_inference/models/jax/gpt_oss.py +26 -10
- tpu_inference/models/jax/jax_intermediate_tensor.py +14 -0
- tpu_inference/models/jax/llama3.py +99 -36
- tpu_inference/models/jax/llama4.py +14 -0
- tpu_inference/models/jax/llama_eagle3.py +14 -0
- tpu_inference/models/jax/llama_guard_4.py +15 -1
- tpu_inference/models/jax/qwen2.py +17 -2
- tpu_inference/models/jax/qwen2_5_vl.py +18 -4
- tpu_inference/models/jax/qwen3.py +17 -2
- tpu_inference/models/jax/utils/__init__.py +13 -0
- tpu_inference/models/jax/utils/file_utils.py +14 -0
- tpu_inference/models/jax/utils/multi_modal_utils.py +18 -4
- tpu_inference/models/jax/utils/qwix/__init__.py +13 -0
- tpu_inference/models/jax/utils/{quantization/quantization_utils.py → qwix/qwix_utils.py} +88 -25
- tpu_inference/models/jax/utils/weight_utils.py +39 -2
- tpu_inference/models/vllm/__init__.py +13 -0
- tpu_inference/models/vllm/vllm_model_wrapper.py +20 -3
- tpu_inference/models/vllm/vllm_model_wrapper_context.py +14 -0
- tpu_inference/platforms/__init__.py +14 -0
- tpu_inference/platforms/tpu_platform.py +47 -64
- tpu_inference/runner/__init__.py +13 -0
- tpu_inference/runner/compilation_manager.py +72 -37
- tpu_inference/runner/kv_cache.py +54 -20
- tpu_inference/runner/kv_cache_manager.py +45 -15
- tpu_inference/runner/lora_utils.py +14 -0
- tpu_inference/runner/multimodal_manager.py +15 -1
- tpu_inference/runner/persistent_batch_manager.py +14 -0
- tpu_inference/runner/speculative_decoding_manager.py +14 -0
- tpu_inference/runner/structured_decoding_manager.py +14 -0
- tpu_inference/runner/tpu_runner.py +41 -16
- tpu_inference/spec_decode/__init__.py +13 -0
- tpu_inference/spec_decode/jax/__init__.py +13 -0
- tpu_inference/spec_decode/jax/eagle3.py +13 -0
- tpu_inference/tpu_info.py +14 -0
- tpu_inference/utils.py +42 -36
- tpu_inference/worker/__init__.py +13 -0
- tpu_inference/worker/tpu_worker.py +63 -50
- {tpu_inference-0.11.1.dev202512030818.dist-info → tpu_inference-0.13.0rc2.post7.dist-info}/METADATA +11 -9
- tpu_inference-0.13.0rc2.post7.dist-info/RECORD +261 -0
- tpu_inference/models/jax/utils/quantization/__init__.py +0 -0
- tpu_inference/models/jax/utils/quantization/configs/fp8_all_modules_w_only.yaml +0 -5
- tpu_inference/models/jax/utils/quantization/configs/fp8_default.yaml +0 -6
- tpu_inference/models/jax/utils/quantization/configs/int8_all_modules_w_only.yaml +0 -5
- tpu_inference/models/jax/utils/quantization/configs/int8_default.yaml +0 -6
- tpu_inference/models/jax/utils/quantization/mxfp4_utils.py +0 -105
- tpu_inference-0.11.1.dev202512030818.dist-info/RECORD +0 -174
- {tpu_inference-0.11.1.dev202512030818.dist-info → tpu_inference-0.13.0rc2.post7.dist-info}/WHEEL +0 -0
- {tpu_inference-0.11.1.dev202512030818.dist-info → tpu_inference-0.13.0rc2.post7.dist-info}/licenses/LICENSE +0 -0
- {tpu_inference-0.11.1.dev202512030818.dist-info → tpu_inference-0.13.0rc2.post7.dist-info}/top_level.txt +0 -0
|
@@ -1,3 +1,16 @@
|
|
|
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.
|
|
1
14
|
"""Auto-tuned block sizes for ragged paged attention."""
|
|
2
15
|
|
|
3
16
|
import jax.numpy as jnp
|
|
@@ -14,240 +27,396 @@ logger = init_logger(__name__)
|
|
|
14
27
|
# - page_size
|
|
15
28
|
# - q_{q_dtype_name}_kv_{kv_dtype_name}
|
|
16
29
|
# - q_head-{num_q_heads}_kv_head-{num_kv_heads}-_head-{head_dim}
|
|
17
|
-
# - max_model_len
|
|
30
|
+
# - max_model_len-{max_model_len}-sw-{sliding_window}
|
|
18
31
|
# value:
|
|
19
32
|
# - (num_kv_pages_per_block, num_queries_per_block)
|
|
20
33
|
TUNED_BLOCK_SIZES = {
|
|
21
|
-
'TPU
|
|
22
|
-
|
|
34
|
+
'TPU v7': {
|
|
35
|
+
256: {
|
|
23
36
|
'q_bfloat16_kv_bfloat16': {
|
|
24
|
-
'q_head-
|
|
25
|
-
|
|
26
|
-
|
|
27
|
-
|
|
28
|
-
|
|
29
|
-
|
|
30
|
-
|
|
31
|
-
|
|
37
|
+
'q_head-16_kv_head-2_head-64': {
|
|
38
|
+
'max_model_len-2048-sw-None': (8, 32),
|
|
39
|
+
'max_model_len-2048-sw-128': (1, 32),
|
|
40
|
+
'max_model_len-4096-sw-None': (16, 32),
|
|
41
|
+
'max_model_len-4096-sw-128': (1, 32),
|
|
42
|
+
'max_model_len-8192-sw-None': (16, 32),
|
|
43
|
+
'max_model_len-8192-sw-128': (1, 32),
|
|
44
|
+
'max_model_len-256-sw-None': (1, 32),
|
|
45
|
+
'max_model_len-256-sw-128': (1, 32),
|
|
46
|
+
'max_model_len-512-sw-None': (2, 32),
|
|
47
|
+
'max_model_len-512-sw-128': (1, 32),
|
|
48
|
+
'max_model_len-1024-sw-None': (4, 32),
|
|
49
|
+
'max_model_len-1024-sw-128': (1, 64),
|
|
32
50
|
},
|
|
33
51
|
'q_head-64_kv_head-8_head-64': {
|
|
34
|
-
128: (1, 16),
|
|
35
|
-
|
|
36
|
-
|
|
37
|
-
256: (
|
|
38
|
-
|
|
39
|
-
|
|
40
|
-
|
|
52
|
+
'max_model_len-8192-sw-128': (1, 16),
|
|
53
|
+
'max_model_len-512-sw-None': (2, 16),
|
|
54
|
+
'max_model_len-256-sw-None': (1, 16),
|
|
55
|
+
'max_model_len-256-sw-128': (1, 16),
|
|
56
|
+
'max_model_len-512-sw-128': (1, 16),
|
|
57
|
+
'max_model_len-1024-sw-None': (4, 16),
|
|
58
|
+
'max_model_len-1024-sw-128': (1, 16),
|
|
59
|
+
'max_model_len-2048-sw-None': (8, 32),
|
|
60
|
+
'max_model_len-2048-sw-128': (1, 16),
|
|
61
|
+
'max_model_len-4096-sw-None': (16, 16),
|
|
62
|
+
'max_model_len-4096-sw-128': (1, 16),
|
|
63
|
+
'max_model_len-8192-sw-None': (16, 16),
|
|
41
64
|
},
|
|
42
65
|
'q_head-32_kv_head-4_head-64': {
|
|
43
|
-
256: (
|
|
44
|
-
|
|
45
|
-
|
|
46
|
-
|
|
47
|
-
4096: (
|
|
48
|
-
|
|
49
|
-
128: (1,
|
|
66
|
+
'max_model_len-256-sw-None': (1, 32),
|
|
67
|
+
'max_model_len-256-sw-128': (1, 32),
|
|
68
|
+
'max_model_len-512-sw-None': (2, 32),
|
|
69
|
+
'max_model_len-512-sw-128': (1, 32),
|
|
70
|
+
'max_model_len-4096-sw-None': (16, 32),
|
|
71
|
+
'max_model_len-1024-sw-None': (4, 32),
|
|
72
|
+
'max_model_len-1024-sw-128': (1, 32),
|
|
73
|
+
'max_model_len-2048-sw-None': (8, 32),
|
|
74
|
+
'max_model_len-2048-sw-128': (1, 32),
|
|
75
|
+
'max_model_len-4096-sw-128': (1, 32),
|
|
76
|
+
'max_model_len-8192-sw-None': (16, 32),
|
|
77
|
+
'max_model_len-8192-sw-128': (1, 32),
|
|
50
78
|
},
|
|
51
|
-
'q_head-
|
|
52
|
-
|
|
53
|
-
256: (
|
|
54
|
-
512: (
|
|
55
|
-
|
|
56
|
-
|
|
57
|
-
|
|
58
|
-
|
|
79
|
+
'q_head-8_kv_head-2_head-64': {
|
|
80
|
+
'max_model_len-256-sw-None': (1, 64),
|
|
81
|
+
'max_model_len-256-sw-128': (1, 64),
|
|
82
|
+
'max_model_len-512-sw-None': (2, 64),
|
|
83
|
+
'max_model_len-512-sw-128': (1, 64),
|
|
84
|
+
'max_model_len-1024-sw-None': (4, 64),
|
|
85
|
+
'max_model_len-1024-sw-128': (1, 64),
|
|
86
|
+
'max_model_len-2048-sw-None': (8, 32),
|
|
87
|
+
'max_model_len-2048-sw-128': (1, 128),
|
|
88
|
+
'max_model_len-4096-sw-None': (16, 32),
|
|
89
|
+
'max_model_len-4096-sw-128': (1, 128),
|
|
90
|
+
'max_model_len-8192-sw-None': (16, 32),
|
|
91
|
+
'max_model_len-8192-sw-128': (1, 64),
|
|
59
92
|
},
|
|
60
93
|
}
|
|
61
94
|
},
|
|
62
|
-
|
|
95
|
+
128: {
|
|
63
96
|
'q_bfloat16_kv_bfloat16': {
|
|
64
|
-
'q_head-16_kv_head-2_head-64': {
|
|
65
|
-
1024: (4, 32),
|
|
66
|
-
2048: (8, 16),
|
|
67
|
-
4096: (8, 32),
|
|
68
|
-
8192: (16, 16),
|
|
69
|
-
256: (1, 128),
|
|
70
|
-
512: (2, 128),
|
|
71
|
-
},
|
|
72
97
|
'q_head-64_kv_head-8_head-64': {
|
|
73
|
-
|
|
74
|
-
|
|
75
|
-
1024: (
|
|
76
|
-
|
|
77
|
-
|
|
78
|
-
|
|
98
|
+
'max_model_len-4096-sw-128': (2, 16),
|
|
99
|
+
'max_model_len-8192-sw-None': (32, 16),
|
|
100
|
+
'max_model_len-1024-sw-None': (8, 16),
|
|
101
|
+
'max_model_len-8192-sw-128': (2, 16),
|
|
102
|
+
'max_model_len-512-sw-None': (4, 16),
|
|
103
|
+
'max_model_len-2048-sw-128': (2, 16),
|
|
104
|
+
'max_model_len-256-sw-None': (2, 16),
|
|
105
|
+
'max_model_len-128-sw-None': (1, 32),
|
|
106
|
+
'max_model_len-4096-sw-None': (32, 16),
|
|
107
|
+
'max_model_len-1024-sw-128': (2, 16),
|
|
108
|
+
'max_model_len-512-sw-128': (2, 16),
|
|
109
|
+
'max_model_len-256-sw-128': (2, 16),
|
|
110
|
+
'max_model_len-2048-sw-None': (16, 32),
|
|
111
|
+
'max_model_len-128-sw-128': (1, 32),
|
|
112
|
+
},
|
|
113
|
+
'q_head-16_kv_head-2_head-64': {
|
|
114
|
+
'max_model_len-128-sw-128': (1, 32),
|
|
115
|
+
'max_model_len-256-sw-None': (2, 32),
|
|
116
|
+
'max_model_len-256-sw-128': (2, 32),
|
|
117
|
+
'max_model_len-512-sw-None': (4, 32),
|
|
118
|
+
'max_model_len-512-sw-128': (2, 32),
|
|
119
|
+
'max_model_len-1024-sw-None': (8, 32),
|
|
120
|
+
'max_model_len-1024-sw-128': (2, 64),
|
|
121
|
+
'max_model_len-2048-sw-None': (16, 32),
|
|
122
|
+
'max_model_len-2048-sw-128': (2, 32),
|
|
123
|
+
'max_model_len-4096-sw-None': (32, 32),
|
|
124
|
+
'max_model_len-4096-sw-128': (2, 32),
|
|
125
|
+
'max_model_len-8192-sw-None': (32, 32),
|
|
126
|
+
'max_model_len-8192-sw-128': (2, 64),
|
|
127
|
+
'max_model_len-128-sw-None': (1, 64),
|
|
79
128
|
},
|
|
80
129
|
'q_head-8_kv_head-2_head-64': {
|
|
81
|
-
|
|
82
|
-
|
|
83
|
-
|
|
84
|
-
|
|
85
|
-
4096: (
|
|
86
|
-
8192: (
|
|
130
|
+
'max_model_len-1024-sw-128': (2, 64),
|
|
131
|
+
'max_model_len-2048-sw-None': (16, 32),
|
|
132
|
+
'max_model_len-2048-sw-128': (2, 32),
|
|
133
|
+
'max_model_len-4096-sw-None': (32, 32),
|
|
134
|
+
'max_model_len-4096-sw-128': (2, 128),
|
|
135
|
+
'max_model_len-8192-sw-None': (32, 32),
|
|
136
|
+
'max_model_len-8192-sw-128': (2, 64),
|
|
137
|
+
'max_model_len-128-sw-None': (1, 128),
|
|
138
|
+
'max_model_len-128-sw-128': (1, 64),
|
|
139
|
+
'max_model_len-256-sw-None': (2, 64),
|
|
140
|
+
'max_model_len-256-sw-128': (2, 32),
|
|
141
|
+
'max_model_len-512-sw-None': (4, 64),
|
|
142
|
+
'max_model_len-512-sw-128': (2, 64),
|
|
143
|
+
'max_model_len-1024-sw-None': (8, 32),
|
|
87
144
|
},
|
|
88
145
|
'q_head-32_kv_head-4_head-64': {
|
|
89
|
-
|
|
90
|
-
|
|
91
|
-
|
|
92
|
-
|
|
93
|
-
|
|
94
|
-
|
|
146
|
+
'max_model_len-128-sw-None': (1, 32),
|
|
147
|
+
'max_model_len-128-sw-128': (1, 32),
|
|
148
|
+
'max_model_len-256-sw-None': (2, 32),
|
|
149
|
+
'max_model_len-256-sw-128': (2, 32),
|
|
150
|
+
'max_model_len-2048-sw-128': (2, 32),
|
|
151
|
+
'max_model_len-512-sw-None': (4, 32),
|
|
152
|
+
'max_model_len-512-sw-128': (2, 32),
|
|
153
|
+
'max_model_len-1024-sw-None': (8, 32),
|
|
154
|
+
'max_model_len-1024-sw-128': (2, 32),
|
|
155
|
+
'max_model_len-2048-sw-None': (16, 32),
|
|
156
|
+
'max_model_len-4096-sw-None': (32, 32),
|
|
157
|
+
'max_model_len-4096-sw-128': (2, 32),
|
|
158
|
+
'max_model_len-8192-sw-None': (32, 32),
|
|
159
|
+
'max_model_len-8192-sw-128': (2, 32),
|
|
95
160
|
},
|
|
96
161
|
}
|
|
97
162
|
},
|
|
98
163
|
},
|
|
99
164
|
'TPU v6e': {
|
|
100
|
-
|
|
165
|
+
256: {
|
|
101
166
|
'q_bfloat16_kv_bfloat16': {
|
|
102
|
-
'q_head-
|
|
103
|
-
|
|
104
|
-
|
|
105
|
-
|
|
106
|
-
|
|
107
|
-
|
|
108
|
-
|
|
109
|
-
|
|
167
|
+
'q_head-16_kv_head-2_head-64': {
|
|
168
|
+
'max_model_len-2048-sw-None': (8, 32),
|
|
169
|
+
'max_model_len-2048-sw-128': (1, 32),
|
|
170
|
+
'max_model_len-4096-sw-None': (16, 32),
|
|
171
|
+
'max_model_len-4096-sw-128': (1, 32),
|
|
172
|
+
'max_model_len-8192-sw-None': (16, 32),
|
|
173
|
+
'max_model_len-8192-sw-128': (1, 32),
|
|
174
|
+
'max_model_len-256-sw-None': (1, 32),
|
|
175
|
+
'max_model_len-256-sw-128': (1, 32),
|
|
176
|
+
'max_model_len-512-sw-None': (2, 32),
|
|
177
|
+
'max_model_len-512-sw-128': (1, 32),
|
|
178
|
+
'max_model_len-1024-sw-None': (4, 32),
|
|
179
|
+
'max_model_len-1024-sw-128': (1, 32),
|
|
110
180
|
},
|
|
111
181
|
'q_head-64_kv_head-8_head-64': {
|
|
112
|
-
128: (1,
|
|
113
|
-
|
|
114
|
-
|
|
115
|
-
256: (
|
|
116
|
-
|
|
117
|
-
|
|
118
|
-
|
|
182
|
+
'max_model_len-8192-sw-128': (1, 16),
|
|
183
|
+
'max_model_len-512-sw-None': (2, 16),
|
|
184
|
+
'max_model_len-256-sw-None': (1, 16),
|
|
185
|
+
'max_model_len-256-sw-128': (1, 16),
|
|
186
|
+
'max_model_len-512-sw-128': (1, 16),
|
|
187
|
+
'max_model_len-1024-sw-None': (4, 16),
|
|
188
|
+
'max_model_len-1024-sw-128': (1, 16),
|
|
189
|
+
'max_model_len-2048-sw-None': (8, 32),
|
|
190
|
+
'max_model_len-2048-sw-128': (1, 16),
|
|
191
|
+
'max_model_len-4096-sw-None': (16, 16),
|
|
192
|
+
'max_model_len-4096-sw-128': (1, 16),
|
|
193
|
+
'max_model_len-8192-sw-None': (16, 16),
|
|
119
194
|
},
|
|
120
195
|
'q_head-32_kv_head-4_head-64': {
|
|
121
|
-
256: (
|
|
122
|
-
|
|
123
|
-
|
|
124
|
-
|
|
125
|
-
4096: (16,
|
|
126
|
-
|
|
127
|
-
128: (1,
|
|
196
|
+
'max_model_len-256-sw-None': (1, 32),
|
|
197
|
+
'max_model_len-256-sw-128': (1, 32),
|
|
198
|
+
'max_model_len-512-sw-None': (2, 32),
|
|
199
|
+
'max_model_len-512-sw-128': (1, 32),
|
|
200
|
+
'max_model_len-4096-sw-None': (16, 32),
|
|
201
|
+
'max_model_len-1024-sw-None': (4, 32),
|
|
202
|
+
'max_model_len-1024-sw-128': (1, 32),
|
|
203
|
+
'max_model_len-2048-sw-None': (8, 32),
|
|
204
|
+
'max_model_len-2048-sw-128': (1, 32),
|
|
205
|
+
'max_model_len-4096-sw-128': (1, 32),
|
|
206
|
+
'max_model_len-8192-sw-None': (16, 32),
|
|
207
|
+
'max_model_len-8192-sw-128': (1, 32),
|
|
128
208
|
},
|
|
129
|
-
'q_head-
|
|
130
|
-
|
|
131
|
-
256: (
|
|
132
|
-
512: (
|
|
133
|
-
|
|
134
|
-
|
|
135
|
-
|
|
136
|
-
|
|
209
|
+
'q_head-8_kv_head-2_head-64': {
|
|
210
|
+
'max_model_len-256-sw-None': (1, 64),
|
|
211
|
+
'max_model_len-256-sw-128': (1, 64),
|
|
212
|
+
'max_model_len-512-sw-None': (2, 64),
|
|
213
|
+
'max_model_len-512-sw-128': (1, 64),
|
|
214
|
+
'max_model_len-1024-sw-None': (4, 32),
|
|
215
|
+
'max_model_len-1024-sw-128': (1, 64),
|
|
216
|
+
'max_model_len-2048-sw-None': (8, 32),
|
|
217
|
+
'max_model_len-2048-sw-128': (1, 64),
|
|
218
|
+
'max_model_len-4096-sw-None': (16, 32),
|
|
219
|
+
'max_model_len-4096-sw-128': (1, 32),
|
|
220
|
+
'max_model_len-8192-sw-None': (16, 32),
|
|
221
|
+
'max_model_len-8192-sw-128': (1, 32),
|
|
137
222
|
},
|
|
138
223
|
}
|
|
139
224
|
},
|
|
140
|
-
|
|
225
|
+
128: {
|
|
141
226
|
'q_bfloat16_kv_bfloat16': {
|
|
142
|
-
'q_head-16_kv_head-2_head-64': {
|
|
143
|
-
1024: (4, 128),
|
|
144
|
-
2048: (8, 32),
|
|
145
|
-
4096: (16, 16),
|
|
146
|
-
8192: (16, 16),
|
|
147
|
-
256: (1, 64),
|
|
148
|
-
512: (2, 32),
|
|
149
|
-
},
|
|
150
227
|
'q_head-64_kv_head-8_head-64': {
|
|
151
|
-
|
|
152
|
-
|
|
153
|
-
1024: (
|
|
154
|
-
|
|
155
|
-
|
|
156
|
-
|
|
228
|
+
'max_model_len-4096-sw-128': (2, 16),
|
|
229
|
+
'max_model_len-8192-sw-None': (16, 32),
|
|
230
|
+
'max_model_len-1024-sw-None': (8, 16),
|
|
231
|
+
'max_model_len-8192-sw-128': (2, 16),
|
|
232
|
+
'max_model_len-512-sw-None': (4, 16),
|
|
233
|
+
'max_model_len-2048-sw-128': (2, 16),
|
|
234
|
+
'max_model_len-256-sw-None': (2, 16),
|
|
235
|
+
'max_model_len-128-sw-None': (1, 32),
|
|
236
|
+
'max_model_len-4096-sw-None': (32, 16),
|
|
237
|
+
'max_model_len-1024-sw-128': (2, 16),
|
|
238
|
+
'max_model_len-512-sw-128': (2, 16),
|
|
239
|
+
'max_model_len-256-sw-128': (2, 16),
|
|
240
|
+
'max_model_len-2048-sw-None': (16, 32),
|
|
241
|
+
'max_model_len-128-sw-128': (1, 32),
|
|
242
|
+
},
|
|
243
|
+
'q_head-16_kv_head-2_head-64': {
|
|
244
|
+
'max_model_len-128-sw-128': (1, 32),
|
|
245
|
+
'max_model_len-256-sw-None': (2, 32),
|
|
246
|
+
'max_model_len-256-sw-128': (2, 32),
|
|
247
|
+
'max_model_len-512-sw-None': (4, 32),
|
|
248
|
+
'max_model_len-512-sw-128': (2, 32),
|
|
249
|
+
'max_model_len-1024-sw-None': (8, 32),
|
|
250
|
+
'max_model_len-1024-sw-128': (2, 32),
|
|
251
|
+
'max_model_len-2048-sw-None': (16, 32),
|
|
252
|
+
'max_model_len-2048-sw-128': (2, 32),
|
|
253
|
+
'max_model_len-4096-sw-None': (32, 32),
|
|
254
|
+
'max_model_len-4096-sw-128': (2, 32),
|
|
255
|
+
'max_model_len-8192-sw-None': (32, 32),
|
|
256
|
+
'max_model_len-8192-sw-128': (2, 32),
|
|
257
|
+
'max_model_len-128-sw-None': (1, 32),
|
|
157
258
|
},
|
|
158
259
|
'q_head-8_kv_head-2_head-64': {
|
|
159
|
-
|
|
160
|
-
|
|
161
|
-
|
|
162
|
-
|
|
163
|
-
4096: (
|
|
164
|
-
8192: (
|
|
260
|
+
'max_model_len-1024-sw-128': (2, 32),
|
|
261
|
+
'max_model_len-2048-sw-None': (16, 32),
|
|
262
|
+
'max_model_len-2048-sw-128': (2, 64),
|
|
263
|
+
'max_model_len-4096-sw-None': (32, 32),
|
|
264
|
+
'max_model_len-4096-sw-128': (2, 32),
|
|
265
|
+
'max_model_len-8192-sw-None': (32, 32),
|
|
266
|
+
'max_model_len-8192-sw-128': (2, 64),
|
|
267
|
+
'max_model_len-128-sw-None': (1, 64),
|
|
268
|
+
'max_model_len-128-sw-128': (1, 64),
|
|
269
|
+
'max_model_len-256-sw-None': (2, 32),
|
|
270
|
+
'max_model_len-256-sw-128': (2, 32),
|
|
271
|
+
'max_model_len-512-sw-None': (4, 32),
|
|
272
|
+
'max_model_len-512-sw-128': (2, 64),
|
|
273
|
+
'max_model_len-1024-sw-None': (8, 32),
|
|
165
274
|
},
|
|
166
275
|
'q_head-32_kv_head-4_head-64': {
|
|
167
|
-
|
|
168
|
-
|
|
169
|
-
|
|
170
|
-
|
|
171
|
-
|
|
172
|
-
|
|
276
|
+
'max_model_len-128-sw-None': (1, 32),
|
|
277
|
+
'max_model_len-128-sw-128': (1, 32),
|
|
278
|
+
'max_model_len-256-sw-None': (2, 32),
|
|
279
|
+
'max_model_len-256-sw-128': (2, 32),
|
|
280
|
+
'max_model_len-2048-sw-128': (2, 32),
|
|
281
|
+
'max_model_len-512-sw-None': (4, 32),
|
|
282
|
+
'max_model_len-512-sw-128': (2, 32),
|
|
283
|
+
'max_model_len-1024-sw-None': (8, 32),
|
|
284
|
+
'max_model_len-1024-sw-128': (2, 32),
|
|
285
|
+
'max_model_len-2048-sw-None': (16, 32),
|
|
286
|
+
'max_model_len-4096-sw-None': (32, 32),
|
|
287
|
+
'max_model_len-4096-sw-128': (2, 32),
|
|
288
|
+
'max_model_len-8192-sw-None': (32, 32),
|
|
289
|
+
'max_model_len-8192-sw-128': (2, 32),
|
|
173
290
|
},
|
|
174
291
|
}
|
|
175
292
|
},
|
|
176
293
|
},
|
|
177
|
-
'TPU
|
|
178
|
-
|
|
294
|
+
'TPU v5e': {
|
|
295
|
+
256: {
|
|
179
296
|
'q_bfloat16_kv_bfloat16': {
|
|
180
|
-
'q_head-
|
|
181
|
-
|
|
182
|
-
|
|
183
|
-
|
|
184
|
-
|
|
185
|
-
|
|
186
|
-
|
|
187
|
-
|
|
297
|
+
'q_head-16_kv_head-2_head-64': {
|
|
298
|
+
'max_model_len-2048-sw-None': (8, 32),
|
|
299
|
+
'max_model_len-2048-sw-128': (1, 32),
|
|
300
|
+
'max_model_len-4096-sw-None': (16, 32),
|
|
301
|
+
'max_model_len-4096-sw-128': (1, 32),
|
|
302
|
+
'max_model_len-8192-sw-None': (16, 32),
|
|
303
|
+
'max_model_len-8192-sw-128': (1, 32),
|
|
304
|
+
'max_model_len-256-sw-None': (1, 32),
|
|
305
|
+
'max_model_len-256-sw-128': (1, 32),
|
|
306
|
+
'max_model_len-512-sw-None': (2, 32),
|
|
307
|
+
'max_model_len-512-sw-128': (1, 32),
|
|
308
|
+
'max_model_len-1024-sw-None': (4, 32),
|
|
309
|
+
'max_model_len-1024-sw-128': (1, 32),
|
|
188
310
|
},
|
|
189
311
|
'q_head-64_kv_head-8_head-64': {
|
|
190
|
-
128: (1,
|
|
191
|
-
|
|
192
|
-
|
|
193
|
-
256: (
|
|
194
|
-
|
|
195
|
-
|
|
196
|
-
|
|
312
|
+
'max_model_len-8192-sw-128': (1, 8),
|
|
313
|
+
'max_model_len-512-sw-None': (2, 16),
|
|
314
|
+
'max_model_len-256-sw-None': (1, 8),
|
|
315
|
+
'max_model_len-256-sw-128': (1, 8),
|
|
316
|
+
'max_model_len-512-sw-128': (1, 8),
|
|
317
|
+
'max_model_len-1024-sw-None': (4, 32),
|
|
318
|
+
'max_model_len-1024-sw-128': (1, 8),
|
|
319
|
+
'max_model_len-2048-sw-None': (4, 32),
|
|
320
|
+
'max_model_len-2048-sw-128': (1, 8),
|
|
321
|
+
'max_model_len-4096-sw-None': (4, 32),
|
|
322
|
+
'max_model_len-4096-sw-128': (1, 8),
|
|
323
|
+
'max_model_len-8192-sw-None': (4, 32),
|
|
197
324
|
},
|
|
198
325
|
'q_head-32_kv_head-4_head-64': {
|
|
199
|
-
256: (
|
|
200
|
-
|
|
201
|
-
|
|
202
|
-
|
|
203
|
-
4096: (
|
|
204
|
-
|
|
205
|
-
128: (1, 16),
|
|
326
|
+
'max_model_len-256-sw-None': (1, 16),
|
|
327
|
+
'max_model_len-256-sw-128': (1, 16),
|
|
328
|
+
'max_model_len-512-sw-None': (2, 16),
|
|
329
|
+
'max_model_len-512-sw-128': (1, 16),
|
|
330
|
+
'max_model_len-4096-sw-None': (8, 32),
|
|
331
|
+
'max_model_len-1024-sw-None': (4, 16),
|
|
332
|
+
'max_model_len-1024-sw-128': (1, 16),
|
|
333
|
+
'max_model_len-2048-sw-None': (8, 32),
|
|
334
|
+
'max_model_len-2048-sw-128': (1, 16),
|
|
335
|
+
'max_model_len-4096-sw-128': (1, 16),
|
|
336
|
+
'max_model_len-8192-sw-None': (8, 32),
|
|
337
|
+
'max_model_len-8192-sw-128': (1, 16),
|
|
206
338
|
},
|
|
207
|
-
'q_head-
|
|
208
|
-
|
|
209
|
-
256: (
|
|
210
|
-
512: (
|
|
211
|
-
|
|
212
|
-
|
|
213
|
-
|
|
214
|
-
|
|
339
|
+
'q_head-8_kv_head-2_head-64': {
|
|
340
|
+
'max_model_len-256-sw-None': (1, 32),
|
|
341
|
+
'max_model_len-256-sw-128': (1, 32),
|
|
342
|
+
'max_model_len-512-sw-None': (2, 32),
|
|
343
|
+
'max_model_len-512-sw-128': (1, 32),
|
|
344
|
+
'max_model_len-1024-sw-None': (4, 32),
|
|
345
|
+
'max_model_len-1024-sw-128': (1, 32),
|
|
346
|
+
'max_model_len-2048-sw-None': (8, 32),
|
|
347
|
+
'max_model_len-2048-sw-128': (1, 32),
|
|
348
|
+
'max_model_len-4096-sw-None': (16, 32),
|
|
349
|
+
'max_model_len-4096-sw-128': (1, 32),
|
|
350
|
+
'max_model_len-8192-sw-None': (8, 32),
|
|
351
|
+
'max_model_len-8192-sw-128': (1, 32),
|
|
215
352
|
},
|
|
216
353
|
}
|
|
217
354
|
},
|
|
218
|
-
|
|
355
|
+
128: {
|
|
219
356
|
'q_bfloat16_kv_bfloat16': {
|
|
220
|
-
'q_head-16_kv_head-2_head-64': {
|
|
221
|
-
1024: (4, 32),
|
|
222
|
-
2048: (8, 16),
|
|
223
|
-
4096: (16, 8),
|
|
224
|
-
8192: (16, 16),
|
|
225
|
-
256: (1, 64),
|
|
226
|
-
512: (2, 32),
|
|
227
|
-
},
|
|
228
357
|
'q_head-64_kv_head-8_head-64': {
|
|
229
|
-
|
|
230
|
-
|
|
231
|
-
1024: (
|
|
232
|
-
|
|
233
|
-
|
|
234
|
-
|
|
358
|
+
'max_model_len-4096-sw-128': (2, 8),
|
|
359
|
+
'max_model_len-8192-sw-None': (8, 32),
|
|
360
|
+
'max_model_len-1024-sw-None': (8, 32),
|
|
361
|
+
'max_model_len-8192-sw-128': (2, 8),
|
|
362
|
+
'max_model_len-512-sw-None': (4, 16),
|
|
363
|
+
'max_model_len-2048-sw-128': (2, 8),
|
|
364
|
+
'max_model_len-256-sw-None': (2, 8),
|
|
365
|
+
'max_model_len-128-sw-None': (1, 8),
|
|
366
|
+
'max_model_len-4096-sw-None': (8, 32),
|
|
367
|
+
'max_model_len-1024-sw-128': (2, 16),
|
|
368
|
+
'max_model_len-512-sw-128': (2, 8),
|
|
369
|
+
'max_model_len-256-sw-128': (2, 8),
|
|
370
|
+
'max_model_len-2048-sw-None': (8, 32),
|
|
371
|
+
'max_model_len-128-sw-128': (1, 8),
|
|
372
|
+
},
|
|
373
|
+
'q_head-16_kv_head-2_head-64': {
|
|
374
|
+
'max_model_len-128-sw-128': (1, 32),
|
|
375
|
+
'max_model_len-256-sw-None': (2, 32),
|
|
376
|
+
'max_model_len-256-sw-128': (2, 32),
|
|
377
|
+
'max_model_len-512-sw-None': (4, 32),
|
|
378
|
+
'max_model_len-512-sw-128': (2, 32),
|
|
379
|
+
'max_model_len-1024-sw-None': (8, 32),
|
|
380
|
+
'max_model_len-1024-sw-128': (2, 32),
|
|
381
|
+
'max_model_len-2048-sw-None': (16, 32),
|
|
382
|
+
'max_model_len-2048-sw-128': (2, 32),
|
|
383
|
+
'max_model_len-4096-sw-None': (32, 32),
|
|
384
|
+
'max_model_len-4096-sw-128': (2, 32),
|
|
385
|
+
'max_model_len-8192-sw-None': (32, 32),
|
|
386
|
+
'max_model_len-8192-sw-128': (2, 32),
|
|
387
|
+
'max_model_len-128-sw-None': (1, 32),
|
|
235
388
|
},
|
|
236
389
|
'q_head-8_kv_head-2_head-64': {
|
|
237
|
-
|
|
238
|
-
|
|
239
|
-
|
|
240
|
-
|
|
241
|
-
4096: (
|
|
242
|
-
8192: (
|
|
390
|
+
'max_model_len-1024-sw-128': (2, 32),
|
|
391
|
+
'max_model_len-2048-sw-None': (16, 32),
|
|
392
|
+
'max_model_len-2048-sw-128': (2, 32),
|
|
393
|
+
'max_model_len-4096-sw-None': (32, 32),
|
|
394
|
+
'max_model_len-4096-sw-128': (2, 32),
|
|
395
|
+
'max_model_len-8192-sw-None': (32, 32),
|
|
396
|
+
'max_model_len-8192-sw-128': (2, 32),
|
|
397
|
+
'max_model_len-128-sw-None': (1, 32),
|
|
398
|
+
'max_model_len-128-sw-128': (1, 32),
|
|
399
|
+
'max_model_len-256-sw-None': (2, 32),
|
|
400
|
+
'max_model_len-256-sw-128': (2, 32),
|
|
401
|
+
'max_model_len-512-sw-None': (4, 32),
|
|
402
|
+
'max_model_len-512-sw-128': (2, 32),
|
|
403
|
+
'max_model_len-1024-sw-None': (8, 32),
|
|
243
404
|
},
|
|
244
405
|
'q_head-32_kv_head-4_head-64': {
|
|
245
|
-
|
|
246
|
-
|
|
247
|
-
|
|
248
|
-
|
|
249
|
-
|
|
250
|
-
|
|
406
|
+
'max_model_len-128-sw-None': (1, 16),
|
|
407
|
+
'max_model_len-128-sw-128': (1, 16),
|
|
408
|
+
'max_model_len-256-sw-None': (2, 16),
|
|
409
|
+
'max_model_len-256-sw-128': (2, 16),
|
|
410
|
+
'max_model_len-2048-sw-128': (2, 16),
|
|
411
|
+
'max_model_len-512-sw-None': (4, 16),
|
|
412
|
+
'max_model_len-512-sw-128': (2, 16),
|
|
413
|
+
'max_model_len-1024-sw-None': (8, 16),
|
|
414
|
+
'max_model_len-1024-sw-128': (2, 16),
|
|
415
|
+
'max_model_len-2048-sw-None': (16, 32),
|
|
416
|
+
'max_model_len-4096-sw-None': (16, 32),
|
|
417
|
+
'max_model_len-4096-sw-128': (2, 16),
|
|
418
|
+
'max_model_len-8192-sw-None': (16, 32),
|
|
419
|
+
'max_model_len-8192-sw-128': (2, 16),
|
|
251
420
|
},
|
|
252
421
|
}
|
|
253
422
|
},
|
|
@@ -264,22 +433,10 @@ def get_tuned_block_sizes(
|
|
|
264
433
|
page_size,
|
|
265
434
|
max_num_tokens,
|
|
266
435
|
pages_per_seq,
|
|
436
|
+
sliding_window=None,
|
|
267
437
|
) -> tuple[int, int]:
|
|
268
438
|
"""Search tuned values for (num_kv_pages_per_blk, num_queries_per_blk)."""
|
|
269
439
|
|
|
270
|
-
# Set default block sizes for each tpu_version.
|
|
271
|
-
tpu_version = get_tpu_version()
|
|
272
|
-
if tpu_version < 4:
|
|
273
|
-
raise NotImplementedError('TPU version must be 4 or higher.')
|
|
274
|
-
match tpu_version:
|
|
275
|
-
case 4:
|
|
276
|
-
# TPUv4 has much smaller VMEM size so we pick fixed block sizes.
|
|
277
|
-
bkv_p, bq = (512 // page_size, 32)
|
|
278
|
-
case 7:
|
|
279
|
-
bkv_p, bq = (4096 // page_size, 32)
|
|
280
|
-
case _:
|
|
281
|
-
bkv_p, bq = (2048 // page_size, 32)
|
|
282
|
-
|
|
283
440
|
keys = get_lookup_keys(
|
|
284
441
|
page_size,
|
|
285
442
|
q_dtype,
|
|
@@ -288,16 +445,35 @@ def get_tuned_block_sizes(
|
|
|
288
445
|
actual_num_kv_heads,
|
|
289
446
|
head_dim,
|
|
290
447
|
page_size * pages_per_seq,
|
|
448
|
+
sliding_window,
|
|
291
449
|
)
|
|
292
|
-
device, page_size, dtypes, head_dims,
|
|
450
|
+
device, page_size, dtypes, head_dims, extra = keys
|
|
293
451
|
|
|
294
452
|
try:
|
|
295
453
|
bkv_p, bq = TUNED_BLOCK_SIZES[device][page_size][dtypes][head_dims][
|
|
296
|
-
|
|
454
|
+
extra]
|
|
297
455
|
except KeyError:
|
|
298
|
-
|
|
456
|
+
logger.warning_once(
|
|
457
|
+
'Couldn`t find tuned sizes for the RPA v3 kernel with %s', keys)
|
|
458
|
+
# When not available use a sensible default based on TPU version
|
|
459
|
+
# Set default block sizes for each tpu_version.
|
|
460
|
+
tpu_version = get_tpu_version()
|
|
461
|
+
if tpu_version < 4:
|
|
462
|
+
raise NotImplementedError('TPU version must be 4 or higher.')
|
|
463
|
+
match tpu_version:
|
|
464
|
+
case 4:
|
|
465
|
+
# TPUv4 has much smaller VMEM size so we pick fixed block sizes.
|
|
466
|
+
bkv_p, bq = (512 // page_size, 32)
|
|
467
|
+
case 7:
|
|
468
|
+
bkv_p, bq = (4096 // page_size, 32)
|
|
469
|
+
case _:
|
|
470
|
+
bkv_p, bq = (2048 // page_size, 32)
|
|
299
471
|
|
|
300
|
-
|
|
472
|
+
bkv_p, bq = (min(pages_per_seq, bkv_p), min(max_num_tokens, bq))
|
|
473
|
+
|
|
474
|
+
logger.info_once('RPA v3 kernel tuned block sizes for %s: bkv_p=%s, bq=%s',
|
|
475
|
+
keys, bkv_p, bq)
|
|
476
|
+
return bkv_p, bq
|
|
301
477
|
|
|
302
478
|
|
|
303
479
|
def get_lookup_keys(
|
|
@@ -308,6 +484,7 @@ def get_lookup_keys(
|
|
|
308
484
|
num_kv_heads,
|
|
309
485
|
head_dim,
|
|
310
486
|
max_model_len,
|
|
487
|
+
sliding_window,
|
|
311
488
|
):
|
|
312
489
|
"""Get the lookup keys for tuned block sizes."""
|
|
313
490
|
(
|
|
@@ -318,6 +495,7 @@ def get_lookup_keys(
|
|
|
318
495
|
num_kv_heads,
|
|
319
496
|
head_dim,
|
|
320
497
|
max_model_len,
|
|
498
|
+
sliding_window,
|
|
321
499
|
) = get_simplified_raw_key(
|
|
322
500
|
page_size,
|
|
323
501
|
q_dtype,
|
|
@@ -326,6 +504,7 @@ def get_lookup_keys(
|
|
|
326
504
|
num_kv_heads,
|
|
327
505
|
head_dim,
|
|
328
506
|
max_model_len,
|
|
507
|
+
sliding_window,
|
|
329
508
|
)
|
|
330
509
|
|
|
331
510
|
return (
|
|
@@ -333,7 +512,7 @@ def get_lookup_keys(
|
|
|
333
512
|
next_power_of_2(page_size),
|
|
334
513
|
f'q_{q_dtype_name}_kv_{kv_dtype_name}',
|
|
335
514
|
f'q_head-{num_q_heads}_kv_head-{num_kv_heads}_head-{head_dim}',
|
|
336
|
-
next_power_of_2(max_model_len),
|
|
515
|
+
f'max_model_len-{next_power_of_2(max_model_len)}-sw-{sliding_window}',
|
|
337
516
|
)
|
|
338
517
|
|
|
339
518
|
|
|
@@ -345,6 +524,7 @@ def get_simplified_raw_key(
|
|
|
345
524
|
actual_num_kv_heads,
|
|
346
525
|
head_dim,
|
|
347
526
|
max_model_len,
|
|
527
|
+
sliding_window,
|
|
348
528
|
):
|
|
349
529
|
"""Get the simplified key."""
|
|
350
530
|
assert head_dim == 64
|
|
@@ -364,4 +544,5 @@ def get_simplified_raw_key(
|
|
|
364
544
|
next_power_of_2(num_kv_heads),
|
|
365
545
|
head_dim,
|
|
366
546
|
next_power_of_2(max_model_len),
|
|
547
|
+
sliding_window,
|
|
367
548
|
)
|