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.
Files changed (260) hide show
  1. tests/__init__.py +13 -0
  2. tests/core/__init__.py +13 -0
  3. tests/core/test_core_tpu.py +513 -0
  4. tests/core/test_disagg_executor.py +60 -0
  5. tests/core/test_disagg_utils.py +67 -0
  6. tests/core/test_dp_scheduler.py +724 -0
  7. tests/core/test_init.py +63 -0
  8. tests/distributed/__init__.py +13 -0
  9. tests/distributed/test_distributed_utils.py +120 -0
  10. tests/distributed/test_tpu_connector.py +478 -0
  11. tests/e2e/__init__.py +13 -0
  12. tests/e2e/test_async_scheduler.py +211 -0
  13. tests/e2e/test_data_parallel.py +393 -0
  14. tests/e2e/test_local_disagg.py +257 -0
  15. tests/e2e/test_model_loader.py +268 -0
  16. tests/e2e/test_multi_modal_inference.py +111 -0
  17. tests/e2e/test_pipeline_parallel.py +265 -0
  18. tests/e2e/test_runai_model_streamer_loader.py +104 -0
  19. tests/e2e/test_sampling_params.py +269 -0
  20. tests/e2e/test_speculative_decoding.py +291 -0
  21. tests/e2e/test_structured_decoding.py +46 -0
  22. tests/executors/__init__.py +13 -0
  23. tests/executors/test_ray_distributed_executor.py +199 -0
  24. tests/experimental/__init__.py +13 -0
  25. tests/experimental/test_llama3_jax_stashed.py +208 -0
  26. tests/kernels/__init__.py +13 -0
  27. tests/kernels/collectives/__init__.py +13 -0
  28. tests/kernels/collectives/all_gather_matmul_kernel_test.py +69 -0
  29. tests/kernels/fused_moe_v1_test.py +388 -0
  30. tests/kernels/gmm_test.py +205 -0
  31. tests/kernels/mla_v1_test.py +498 -0
  32. tests/kernels/quantized_matmul_kernel_test.py +159 -0
  33. tests/kernels/ragged_kv_cache_update_v2_test.py +248 -0
  34. tests/kernels/ragged_paged_attention_kernel_v2_test.py +414 -0
  35. tests/kernels/ragged_paged_attention_kernel_v3_hd64_test.py +565 -0
  36. tests/kernels/ragged_paged_attention_kernel_v3_test.py +520 -0
  37. tests/layers/__init__.py +13 -0
  38. tests/layers/common/__init__.py +13 -0
  39. tests/layers/common/test_attention_interface.py +156 -0
  40. tests/layers/common/test_quantization.py +149 -0
  41. tests/layers/jax/__init__.py +13 -0
  42. tests/layers/jax/attention/__init__.py +13 -0
  43. tests/layers/jax/attention/test_common_attention.py +103 -0
  44. tests/layers/jax/attention/test_deepseek_v3_attention.py +233 -0
  45. tests/layers/jax/attention/test_llama4_attention.py +135 -0
  46. tests/layers/jax/moe/__init__.py +13 -0
  47. tests/layers/jax/moe/test_deepseek_moe.py +235 -0
  48. tests/layers/jax/sample/__init__.py +13 -0
  49. tests/layers/jax/sample/test_rejection_sampler.py +1624 -0
  50. tests/layers/jax/sample/test_sampling.py +115 -0
  51. tests/layers/jax/sample/test_sampling_metadata.py +254 -0
  52. tests/layers/jax/test_layers.py +155 -0
  53. tests/layers/jax/test_qwix.py +969 -0
  54. tests/layers/jax/test_rope.py +93 -0
  55. tests/layers/jax/test_sharding.py +159 -0
  56. tests/layers/jax/test_transformer_block.py +152 -0
  57. tests/layers/vllm/__init__.py +13 -0
  58. tests/layers/vllm/test_attention.py +363 -0
  59. tests/layers/vllm/test_awq.py +405 -0
  60. tests/layers/vllm/test_compressed_tensors_moe.py +202 -0
  61. tests/layers/vllm/test_compressed_tensors_w8a8_fp8.py +403 -0
  62. tests/layers/vllm/test_compressed_tensors_w8a8_int8.py +426 -0
  63. tests/layers/vllm/test_fp8.py +17 -0
  64. tests/layers/vllm/test_mxfp4.py +297 -0
  65. tests/layers/vllm/test_unquantized.py +621 -0
  66. tests/layers/vllm/utils.py +72 -0
  67. tests/lora/__init__.py +13 -0
  68. tests/lora/conftest.py +46 -0
  69. tests/lora/test_bgmv.py +57 -0
  70. tests/lora/test_layers.py +666 -0
  71. tests/lora/test_lora.py +147 -0
  72. tests/lora/test_lora_perf.py +67 -0
  73. tests/lora/utils.py +88 -0
  74. tests/models/__init__.py +13 -0
  75. tests/models/common/__init__.py +13 -0
  76. tests/models/common/test_model_loader.py +455 -0
  77. tests/models/jax/__init__.py +13 -0
  78. tests/models/jax/test_deepseek_v3.py +401 -0
  79. tests/models/jax/test_llama3.py +184 -0
  80. tests/models/jax/test_llama4.py +298 -0
  81. tests/models/jax/test_llama_eagle3.py +197 -0
  82. tests/models/jax/test_llama_guard_4.py +242 -0
  83. tests/models/jax/test_qwen2.py +172 -0
  84. tests/models/jax/test_qwen2_5_vl.py +606 -0
  85. tests/models/jax/test_qwen3.py +169 -0
  86. tests/models/jax/test_weight_loading.py +180 -0
  87. tests/models/jax/utils/__init__.py +13 -0
  88. tests/models/jax/utils/test_multi_modal_utils.py +212 -0
  89. tests/platforms/__init__.py +13 -0
  90. tests/platforms/test_tpu_platform.py +54 -0
  91. tests/runner/__init__.py +13 -0
  92. tests/runner/test_block_table.py +395 -0
  93. tests/runner/test_input_batch.py +226 -0
  94. tests/runner/test_kv_cache.py +220 -0
  95. tests/runner/test_kv_cache_manager.py +498 -0
  96. tests/runner/test_multimodal_manager.py +429 -0
  97. tests/runner/test_persistent_batch_manager.py +84 -0
  98. tests/runner/test_speculative_decoding_manager.py +368 -0
  99. tests/runner/test_structured_decoding_manager.py +220 -0
  100. tests/runner/test_tpu_runner.py +202 -0
  101. tests/runner/test_tpu_runner_dp.py +1033 -0
  102. tests/runner/test_tpu_runner_mesh.py +200 -0
  103. tests/runner/test_utils.py +411 -0
  104. tests/spec_decode/__init__.py +13 -0
  105. tests/spec_decode/test_eagle3.py +311 -0
  106. tests/test_base.py +215 -0
  107. tests/test_envs.py +280 -0
  108. tests/test_tpu_info.py +134 -0
  109. tests/test_utils.py +193 -0
  110. tests/worker/__init__.py +13 -0
  111. tests/worker/tpu_worker_test.py +414 -0
  112. tpu_inference/__init__.py +67 -0
  113. tpu_inference/core/__init__.py +13 -0
  114. tpu_inference/core/core_tpu.py +786 -0
  115. tpu_inference/core/disagg_executor.py +118 -0
  116. tpu_inference/core/disagg_utils.py +49 -0
  117. tpu_inference/core/sched/__init__.py +13 -0
  118. tpu_inference/core/sched/dp_scheduler.py +814 -0
  119. tpu_inference/distributed/__init__.py +13 -0
  120. tpu_inference/distributed/jax_parallel_state.py +81 -0
  121. tpu_inference/distributed/tpu_connector.py +732 -0
  122. tpu_inference/distributed/utils.py +112 -0
  123. tpu_inference/env_override.py +9 -0
  124. tpu_inference/envs.py +191 -0
  125. tpu_inference/executors/__init__.py +13 -0
  126. tpu_inference/executors/ray_distributed_executor.py +399 -0
  127. tpu_inference/experimental/__init__.py +13 -0
  128. tpu_inference/experimental/llama3_jax_stashed.py +272 -0
  129. tpu_inference/kernels/__init__.py +13 -0
  130. tpu_inference/kernels/collectives/__init__.py +13 -0
  131. tpu_inference/kernels/collectives/all_gather_matmul.py +741 -0
  132. tpu_inference/kernels/collectives/all_gather_matmul_tuned_block_sizes.py +65 -0
  133. tpu_inference/kernels/collectives/util.py +47 -0
  134. tpu_inference/kernels/flash_attention/__init__.py +13 -0
  135. tpu_inference/kernels/flash_attention/kernel.py +772 -0
  136. tpu_inference/kernels/fused_moe/__init__.py +13 -0
  137. tpu_inference/kernels/fused_moe/v1/__init__.py +13 -0
  138. tpu_inference/kernels/fused_moe/v1/kernel.py +1612 -0
  139. tpu_inference/kernels/megablox/__init__.py +13 -0
  140. tpu_inference/kernels/megablox/common.py +54 -0
  141. tpu_inference/kernels/megablox/gmm.py +646 -0
  142. tpu_inference/kernels/mla/__init__.py +13 -0
  143. tpu_inference/kernels/mla/v1/__init__.py +13 -0
  144. tpu_inference/kernels/mla/v1/kernel.py +1340 -0
  145. tpu_inference/kernels/quantized_matmul/__init__.py +13 -0
  146. tpu_inference/kernels/quantized_matmul/kernel.py +456 -0
  147. tpu_inference/kernels/quantized_matmul/tuned_block_sizes.py +609 -0
  148. tpu_inference/kernels/quantized_matmul/util.py +58 -0
  149. tpu_inference/kernels/ragged_paged_attention/__init__.py +13 -0
  150. tpu_inference/kernels/ragged_paged_attention/v2/__init__.py +13 -0
  151. tpu_inference/kernels/ragged_paged_attention/v2/kernel.py +876 -0
  152. tpu_inference/kernels/ragged_paged_attention/v2/ragged_kv_cache_update.py +288 -0
  153. tpu_inference/kernels/ragged_paged_attention/v2/tuned_block_sizes.py +1482 -0
  154. tpu_inference/kernels/ragged_paged_attention/v3/__init__.py +13 -0
  155. tpu_inference/kernels/ragged_paged_attention/v3/kernel.py +1594 -0
  156. tpu_inference/kernels/ragged_paged_attention/v3/kernel_hd64.py +1586 -0
  157. tpu_inference/kernels/ragged_paged_attention/v3/tuned_block_sizes.py +4460 -0
  158. tpu_inference/kernels/ragged_paged_attention/v3/tuned_block_sizes_hd64.py +548 -0
  159. tpu_inference/kernels/ragged_paged_attention/v3/util.py +65 -0
  160. tpu_inference/layers/__init__.py +13 -0
  161. tpu_inference/layers/common/__init__.py +13 -0
  162. tpu_inference/layers/common/attention_interface.py +403 -0
  163. tpu_inference/layers/common/attention_metadata.py +48 -0
  164. tpu_inference/layers/common/binary_search.py +295 -0
  165. tpu_inference/layers/common/quant_methods.py +23 -0
  166. tpu_inference/layers/common/quantization.py +270 -0
  167. tpu_inference/layers/common/sharding.py +600 -0
  168. tpu_inference/layers/jax/__init__.py +13 -0
  169. tpu_inference/layers/jax/attention/__init__.py +13 -0
  170. tpu_inference/layers/jax/attention/attention.py +268 -0
  171. tpu_inference/layers/jax/attention/deepseek_v3_attention.py +547 -0
  172. tpu_inference/layers/jax/attention/gpt_oss_attention.py +275 -0
  173. tpu_inference/layers/jax/attention/llama4_attention.py +167 -0
  174. tpu_inference/layers/jax/base.py +165 -0
  175. tpu_inference/layers/jax/constants.py +101 -0
  176. tpu_inference/layers/jax/layers.py +315 -0
  177. tpu_inference/layers/jax/misc.py +30 -0
  178. tpu_inference/layers/jax/moe/__init__.py +13 -0
  179. tpu_inference/layers/jax/moe/deepseek_v3_moe.py +615 -0
  180. tpu_inference/layers/jax/moe/gpt_oss_moe.py +199 -0
  181. tpu_inference/layers/jax/moe/moe.py +249 -0
  182. tpu_inference/layers/jax/pp_utils.py +53 -0
  183. tpu_inference/layers/jax/rope.py +294 -0
  184. tpu_inference/layers/jax/rope_interface.py +228 -0
  185. tpu_inference/layers/jax/sample/__init__.py +13 -0
  186. tpu_inference/layers/jax/sample/rejection_sampler.py +528 -0
  187. tpu_inference/layers/jax/sample/sampling.py +110 -0
  188. tpu_inference/layers/jax/sample/sampling_metadata.py +90 -0
  189. tpu_inference/layers/jax/transformer_block.py +121 -0
  190. tpu_inference/layers/vllm/__init__.py +13 -0
  191. tpu_inference/layers/vllm/attention.py +221 -0
  192. tpu_inference/layers/vllm/fused_moe.py +502 -0
  193. tpu_inference/layers/vllm/linear_common.py +221 -0
  194. tpu_inference/layers/vllm/quantization/__init__.py +55 -0
  195. tpu_inference/layers/vllm/quantization/awq.py +221 -0
  196. tpu_inference/layers/vllm/quantization/common.py +124 -0
  197. tpu_inference/layers/vllm/quantization/compressed_tensors/__init__.py +13 -0
  198. tpu_inference/layers/vllm/quantization/compressed_tensors/compressed_tensors.py +135 -0
  199. tpu_inference/layers/vllm/quantization/compressed_tensors/compressed_tensors_moe.py +266 -0
  200. tpu_inference/layers/vllm/quantization/compressed_tensors/schemes/__init__.py +13 -0
  201. tpu_inference/layers/vllm/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_fp8.py +222 -0
  202. tpu_inference/layers/vllm/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_int8.py +150 -0
  203. tpu_inference/layers/vllm/quantization/fp8.py +118 -0
  204. tpu_inference/layers/vllm/quantization/mxfp4.py +396 -0
  205. tpu_inference/layers/vllm/quantization/unquantized.py +416 -0
  206. tpu_inference/layers/vllm/sharding.py +244 -0
  207. tpu_inference/logger.py +10 -0
  208. tpu_inference/lora/__init__.py +13 -0
  209. tpu_inference/lora/torch_lora_ops.py +98 -0
  210. tpu_inference/lora/torch_punica_tpu.py +310 -0
  211. tpu_inference/models/__init__.py +13 -0
  212. tpu_inference/models/common/__init__.py +13 -0
  213. tpu_inference/models/common/model_loader.py +520 -0
  214. tpu_inference/models/jax/__init__.py +13 -0
  215. tpu_inference/models/jax/deepseek_v3.py +978 -0
  216. tpu_inference/models/jax/gpt_oss.py +508 -0
  217. tpu_inference/models/jax/jax_intermediate_tensor.py +93 -0
  218. tpu_inference/models/jax/llama3.py +436 -0
  219. tpu_inference/models/jax/llama4.py +643 -0
  220. tpu_inference/models/jax/llama_eagle3.py +350 -0
  221. tpu_inference/models/jax/llama_guard_4.py +375 -0
  222. tpu_inference/models/jax/qwen2.py +390 -0
  223. tpu_inference/models/jax/qwen2_5_vl.py +1232 -0
  224. tpu_inference/models/jax/qwen3.py +318 -0
  225. tpu_inference/models/jax/utils/__init__.py +13 -0
  226. tpu_inference/models/jax/utils/file_utils.py +110 -0
  227. tpu_inference/models/jax/utils/multi_modal_utils.py +177 -0
  228. tpu_inference/models/jax/utils/qwix/__init__.py +13 -0
  229. tpu_inference/models/jax/utils/qwix/qwix_utils.py +713 -0
  230. tpu_inference/models/jax/utils/weight_utils.py +621 -0
  231. tpu_inference/models/vllm/__init__.py +13 -0
  232. tpu_inference/models/vllm/vllm_model_wrapper.py +307 -0
  233. tpu_inference/models/vllm/vllm_model_wrapper_context.py +59 -0
  234. tpu_inference/platforms/__init__.py +16 -0
  235. tpu_inference/platforms/tpu_platform.py +258 -0
  236. tpu_inference/runner/__init__.py +13 -0
  237. tpu_inference/runner/block_table.py +122 -0
  238. tpu_inference/runner/compilation_manager.py +890 -0
  239. tpu_inference/runner/input_batch.py +435 -0
  240. tpu_inference/runner/kv_cache.py +166 -0
  241. tpu_inference/runner/kv_cache_manager.py +508 -0
  242. tpu_inference/runner/lora_utils.py +106 -0
  243. tpu_inference/runner/multimodal_manager.py +231 -0
  244. tpu_inference/runner/persistent_batch_manager.py +296 -0
  245. tpu_inference/runner/speculative_decoding_manager.py +262 -0
  246. tpu_inference/runner/structured_decoding_manager.py +101 -0
  247. tpu_inference/runner/tpu_runner.py +1768 -0
  248. tpu_inference/runner/utils.py +426 -0
  249. tpu_inference/spec_decode/__init__.py +13 -0
  250. tpu_inference/spec_decode/jax/__init__.py +13 -0
  251. tpu_inference/spec_decode/jax/eagle3.py +430 -0
  252. tpu_inference/tpu_info.py +92 -0
  253. tpu_inference/utils.py +345 -0
  254. tpu_inference/worker/__init__.py +13 -0
  255. tpu_inference/worker/tpu_worker.py +468 -0
  256. tpu_inference-0.12.0.dev20251222.dist-info/METADATA +106 -0
  257. tpu_inference-0.12.0.dev20251222.dist-info/RECORD +260 -0
  258. tpu_inference-0.12.0.dev20251222.dist-info/WHEEL +5 -0
  259. tpu_inference-0.12.0.dev20251222.dist-info/licenses/LICENSE +201 -0
  260. tpu_inference-0.12.0.dev20251222.dist-info/top_level.txt +2 -0
@@ -0,0 +1,609 @@
1
+ # SPDX-License-Identifier: Apache-2.0
2
+ """Tuned block sizes for quantized matmul kernel."""
3
+
4
+ import re
5
+ from typing import NamedTuple
6
+
7
+ import jax
8
+
9
+ from tpu_inference.logger import init_logger
10
+
11
+ logger = init_logger(__name__)
12
+
13
+
14
+ class TunedKey(NamedTuple):
15
+ tpu_version: int
16
+ n_batch: int
17
+ n_out: int
18
+ n_in: int
19
+ x_q_dtype: str
20
+ w_q_dtype: str
21
+
22
+
23
+ class TunedValue(NamedTuple):
24
+ batch_block_size: int
25
+ out_block_size: int
26
+ in_block_size: int
27
+
28
+
29
+ TUNED_BLOCK_SIZES_RAW = {
30
+ # go/keep-sorted start
31
+ (6, 1024, 1024, 4096, 'int8', 'int8'): (1024, 256, 4096),
32
+ (6, 1024, 1024, 8192, 'int8', 'int8'): (1024, 512, 8192),
33
+ (6, 1024, 128, 8192, 'int8', 'int8'): (512, 128, 8192),
34
+ (6, 1024, 1280, 8192, 'int8', 'int8'): (1024, 1280, 2048),
35
+ (6, 1024, 13824, 5120, 'int8', 'int8'): (1024, 768, 5120),
36
+ (6, 1024, 14336, 4096, 'int8', 'int8'): (1024, 1024, 4096),
37
+ (6, 1024, 1536, 1536, 'int8', 'int8'): (256, 1536, 1536),
38
+ (6, 1024, 1536, 8960, 'int8', 'int8'): (1024, 256, 8960),
39
+ (6, 1024, 1792, 5120, 'int8', 'int8'): (1024, 256, 5120),
40
+ (6, 1024, 17920, 1536, 'int8', 'int8'): (1024, 1792, 1536),
41
+ (6, 1024, 2048, 1536, 'int8', 'int8'): (256, 2048, 1536),
42
+ (6, 1024, 28672, 4096, 'int8', 'int8'): (1024, 2048, 4096),
43
+ (6, 1024, 3584, 18944, 'int8', 'int8'): (512, 512, 18944),
44
+ (6, 1024, 3584, 3584, 'int8', 'int8'): (1024, 896, 3584),
45
+ (6, 1024, 3584, 8192, 'int8', 'int8'): (1024, 512, 8192),
46
+ (6, 1024, 37888, 3584, 'int8', 'int8'): (1024, 1024, 3584),
47
+ (6, 1024, 4096, 14336, 'int8', 'int8'): (1024, 256, 14336),
48
+ (6, 1024, 4096, 4096, 'int8', 'int8'): (1024, 256, 4096),
49
+ (6, 1024, 4608, 3584, 'int8', 'int8'): (1024, 768, 3584),
50
+ (6, 1024, 5120, 1280, 'int8', 'int8'): (1024, 1280, 1280),
51
+ (6, 1024, 5120, 3456, 'int8', 'int8'): (1024, 512, 3456),
52
+ (6, 1024, 5120, 640, 'int8', 'int8'): (256, 5120, 640),
53
+ (6, 1024, 5120, 6912, 'int8', 'int8'): (1024, 512, 6912),
54
+ (6, 1024, 6144, 4096, 'int8', 'int8'): (1024, 768, 4096),
55
+ (6, 1024, 6912, 5120, 'int8', 'int8'): (1024, 768, 5120),
56
+ (6, 1024, 7168, 8192, 'int8', 'int8'): (1024, 256, 8192),
57
+ (6, 1024, 8192, 1024, 'int8', 'int8'): (1024, 2048, 1024),
58
+ (6, 1024, 8192, 3584, 'int8', 'int8'): (1024, 1024, 3584),
59
+ (6, 1024, 896, 5120, 'int8', 'int8'): (1024, 896, 2560),
60
+ (6, 128, 1024, 4096, 'int8', 'int8'): (128, 512, 4096),
61
+ (6, 128, 1024, 8192, 'int8', 'int8'): (128, 1024, 2048),
62
+ (6, 128, 128, 8192, 'int8', 'int8'): (128, 128, 4096),
63
+ (6, 128, 1280, 8192, 'int8', 'int8'): (128, 1280, 2048),
64
+ (6, 128, 13824, 5120, 'int8', 'int8'): (128, 13824, 256),
65
+ (6, 128, 14336, 4096, 'int8', 'int8'): (128, 896, 4096),
66
+ (6, 128, 1536, 1536, 'int8', 'int8'): (128, 768, 1536),
67
+ (6, 128, 1536, 8960, 'int8', 'int8'): (128, 384, 8960),
68
+ (6, 128, 1792, 5120, 'int8', 'int8'): (128, 1792, 1280),
69
+ (6, 128, 17920, 1536, 'int8', 'int8'): (128, 8960, 384),
70
+ (6, 128, 2048, 1536, 'int8', 'int8'): (128, 1024, 1536),
71
+ (6, 128, 28672, 4096, 'int8', 'int8'): (128, 28672, 256),
72
+ (6, 128, 3584, 18944, 'int8', 'int8'): (128, 512, 9472),
73
+ (6, 128, 3584, 3584, 'int8', 'int8'): (128, 3584, 896),
74
+ (6, 128, 3584, 8192, 'int8', 'int8'): (128, 3584, 1024),
75
+ (6, 128, 37888, 3584, 'int8', 'int8'): (128, 1024, 3584),
76
+ (6, 128, 4096, 14336, 'int8', 'int8'): (128, 256, 14336),
77
+ (6, 128, 4096, 4096, 'int8', 'int8'): (128, 512, 4096),
78
+ (6, 128, 4608, 3584, 'int8', 'int8'): (128, 768, 3584),
79
+ (6, 128, 5120, 1280, 'int8', 'int8'): (128, 2560, 1280),
80
+ (6, 128, 5120, 3456, 'int8', 'int8'): (128, 640, 3456),
81
+ (6, 128, 5120, 640, 'int8', 'int8'): (128, 5120, 128),
82
+ (6, 128, 5120, 6912, 'int8', 'int8'): (128, 512, 6912),
83
+ (6, 128, 6144, 4096, 'int8', 'int8'): (128, 1024, 4096),
84
+ (6, 128, 6912, 5120, 'int8', 'int8'): (128, 768, 5120),
85
+ (6, 128, 7168, 8192, 'int8', 'int8'): (128, 3584, 1024),
86
+ (6, 128, 8192, 1024, 'int8', 'int8'): (128, 4096, 1024),
87
+ (6, 128, 8192, 3584, 'int8', 'int8'): (128, 8192, 512),
88
+ (6, 128, 896, 5120, 'int8', 'int8'): (128, 896, 2560),
89
+ (6, 16, 1024, 4096, 'int8', 'int8'): (16, 1024, 2048),
90
+ (6, 16, 1024, 8192, 'int8', 'int8'): (16, 256, 8192),
91
+ (6, 16, 128, 8192, 'int8', 'int8'): (16, 128, 8192),
92
+ (6, 16, 1280, 8192, 'int8', 'int8'): (16, 256, 8192),
93
+ (6, 16, 13824, 5120, 'int8', 'int8'): (16, 512, 5120),
94
+ (6, 16, 14336, 4096, 'int8', 'int8'): (16, 896, 4096),
95
+ (6, 16, 1536, 1536, 'int8', 'int8'): (16, 768, 1536),
96
+ (6, 16, 1536, 8960, 'int8', 'int8'): (16, 1536, 1792),
97
+ (6, 16, 1792, 5120, 'int8', 'int8'): (16, 1792, 2560),
98
+ (6, 16, 17920, 1536, 'int8', 'int8'): (16, 1792, 1536),
99
+ (6, 16, 2048, 1536, 'int8', 'int8'): (16, 1024, 1536),
100
+ (6, 16, 28672, 4096, 'int8', 'int8'): (16, 1024, 4096),
101
+ (6, 16, 3584, 18944, 'int8', 'int8'): (16, 256, 18944),
102
+ (6, 16, 3584, 3584, 'int8', 'int8'): (16, 896, 3584),
103
+ (6, 16, 3584, 8192, 'int8', 'int8'): (16, 3584, 1024),
104
+ (6, 16, 37888, 3584, 'int8', 'int8'): (16, 1024, 3584),
105
+ (6, 16, 4096, 14336, 'int8', 'int8'): (16, 2048, 1792),
106
+ (6, 16, 4096, 4096, 'int8', 'int8'): (16, 1024, 4096),
107
+ (6, 16, 4608, 3584, 'int8', 'int8'): (16, 1152, 3584),
108
+ (6, 16, 5120, 1280, 'int8', 'int8'): (16, 2560, 1280),
109
+ (6, 16, 5120, 3456, 'int8', 'int8'): (16, 1024, 3456),
110
+ (6, 16, 5120, 640, 'int8', 'int8'): (16, 2560, 640),
111
+ (6, 16, 5120, 6912, 'int8', 'int8'): (16, 640, 6912),
112
+ (6, 16, 6144, 4096, 'int8', 'int8'): (16, 768, 4096),
113
+ (6, 16, 6912, 5120, 'int8', 'int8'): (16, 768, 5120),
114
+ (6, 16, 7168, 8192, 'int8', 'int8'): (16, 512, 8192),
115
+ (6, 16, 8192, 1024, 'int8', 'int8'): (16, 2048, 1024),
116
+ (6, 16, 8192, 3584, 'int8', 'int8'): (16, 1024, 3584),
117
+ (6, 16, 896, 5120, 'int8', 'int8'): (16, 896, 2560),
118
+ (6, 16384, 1024, 4096, 'int8', 'int8'): (2048, 1024, 4096),
119
+ (6, 16384, 1024, 8192, 'int8', 'int8'): (1024, 512, 8192),
120
+ (6, 16384, 128, 8192, 'int8', 'int8'): (256, 128, 8192),
121
+ (6, 16384, 1280, 8192, 'int8', 'int8'): (512, 1280, 8192),
122
+ (6, 16384, 13824, 5120, 'int8', 'int8'): (256, 2304, 5120),
123
+ (6, 16384, 14336, 4096, 'int8', 'int8'): (1024, 2048, 4096),
124
+ (6, 16384, 1536, 1536, 'int8', 'int8'): (1024, 1536, 1536),
125
+ (6, 16384, 1536, 8960, 'int8', 'int8'): (1024, 1536, 8960),
126
+ (6, 16384, 1792, 5120, 'int8', 'int8'): (512, 1792, 5120),
127
+ (6, 16384, 17920, 1536, 'int8', 'int8'): (512, 17920, 1536),
128
+ (6, 16384, 2048, 1536, 'int8', 'int8'): (4096, 2048, 1536),
129
+ (6, 16384, 28672, 4096, 'int8', 'int8'): (2048, 1792, 4096),
130
+ (6, 16384, 3584, 18944, 'int8', 'int8'): (2048, 3584, 512),
131
+ (6, 16384, 3584, 3584, 'int8', 'int8'): (1024, 3584, 3584),
132
+ (6, 16384, 3584, 8192, 'int8', 'int8'): (1024, 1792, 8192),
133
+ (6, 16384, 37888, 3584, 'int8', 'int8'): (2048, 1024, 3584),
134
+ (6, 16384, 4096, 14336, 'int8', 'int8'): (2048, 4096, 512),
135
+ (6, 16384, 4096, 4096, 'int8', 'int8'): (1024, 4096, 4096),
136
+ (6, 16384, 4608, 3584, 'int8', 'int8'): (1024, 2304, 3584),
137
+ (6, 16384, 5120, 1280, 'int8', 'int8'): (1024, 5120, 1280),
138
+ (6, 16384, 5120, 3456, 'int8', 'int8'): (512, 5120, 3456),
139
+ (6, 16384, 5120, 640, 'int8', 'int8'): (1024, 5120, 640),
140
+ (6, 16384, 5120, 6912, 'int8', 'int8'): (512, 5120, 6912),
141
+ (6, 16384, 6144, 4096, 'int8', 'int8'): (1024, 6144, 4096),
142
+ (6, 16384, 6912, 5120, 'int8', 'int8'): (512, 6912, 5120),
143
+ (6, 16384, 7168, 8192, 'int8', 'int8'): (1024, 1792, 4096),
144
+ (6, 16384, 8192, 1024, 'int8', 'int8'): (256, 8192, 1024),
145
+ (6, 16384, 8192, 3584, 'int8', 'int8'): (1024, 8192, 3584),
146
+ (6, 16384, 896, 5120, 'int8', 'int8'): (1024, 896, 2560),
147
+ (6, 2048, 1024, 4096, 'int8', 'int8'): (2048, 256, 4096),
148
+ (6, 2048, 1024, 8192, 'int8', 'int8'): (2048, 256, 8192),
149
+ (6, 2048, 128, 8192, 'int8', 'int8'): (1024, 128, 8192),
150
+ (6, 2048, 1280, 8192, 'int8', 'int8'): (2048, 256, 8192),
151
+ (6, 2048, 13824, 5120, 'int8', 'int8'): (1024, 768, 5120),
152
+ (6, 2048, 14336, 4096, 'int8', 'int8'): (2048, 1024, 4096),
153
+ (6, 2048, 1536, 1536, 'int8', 'int8'): (2048, 512, 1536),
154
+ (6, 2048, 1536, 8960, 'int8', 'int8'): (1024, 768, 8960),
155
+ (6, 2048, 1792, 5120, 'int8', 'int8'): (2048, 256, 5120),
156
+ (6, 2048, 17920, 1536, 'int8', 'int8'): (2048, 2560, 1536),
157
+ (6, 2048, 2048, 1536, 'int8', 'int8'): (256, 2048, 1536),
158
+ (6, 2048, 28672, 4096, 'int8', 'int8'): (2048, 1024, 4096),
159
+ (6, 2048, 3584, 18944, 'int8', 'int8'): (2048, 3584, 512),
160
+ (6, 2048, 3584, 3584, 'int8', 'int8'): (2048, 512, 3584),
161
+ (6, 2048, 3584, 8192, 'int8', 'int8'): (2048, 512, 8192),
162
+ (6, 2048, 37888, 3584, 'int8', 'int8'): (2048, 1024, 3584),
163
+ (6, 2048, 4096, 14336, 'int8', 'int8'): (2048, 4096, 512),
164
+ (6, 2048, 4096, 4096, 'int8', 'int8'): (2048, 512, 4096),
165
+ (6, 2048, 4608, 3584, 'int8', 'int8'): (2048, 512, 3584),
166
+ (6, 2048, 5120, 1280, 'int8', 'int8'): (2048, 1280, 1280),
167
+ (6, 2048, 5120, 3456, 'int8', 'int8'): (2048, 512, 3456),
168
+ (6, 2048, 5120, 640, 'int8', 'int8'): (256, 5120, 640),
169
+ (6, 2048, 5120, 6912, 'int8', 'int8'): (2048, 512, 6912),
170
+ (6, 2048, 6144, 4096, 'int8', 'int8'): (2048, 512, 4096),
171
+ (6, 2048, 6912, 5120, 'int8', 'int8'): (2048, 768, 5120),
172
+ (6, 2048, 7168, 8192, 'int8', 'int8'): (1024, 512, 8192),
173
+ (6, 2048, 8192, 1024, 'int8', 'int8'): (512, 8192, 1024),
174
+ (6, 2048, 8192, 3584, 'int8', 'int8'): (2048, 512, 3584),
175
+ (6, 2048, 896, 5120, 'int8', 'int8'): (512, 896, 5120),
176
+ (6, 256, 1024, 4096, 'int8', 'int8'): (256, 512, 4096),
177
+ (6, 256, 1024, 8192, 'int8', 'int8'): (256, 512, 8192),
178
+ (6, 256, 128, 8192, 'int8', 'int8'): (256, 128, 8192),
179
+ (6, 256, 1280, 8192, 'int8', 'int8'): (256, 1280, 2048),
180
+ (6, 256, 13824, 5120, 'int8', 'int8'): (256, 6912, 512),
181
+ (6, 256, 14336, 4096, 'int8', 'int8'): (256, 1024, 4096),
182
+ (6, 256, 1536, 1536, 'int8', 'int8'): (256, 768, 1536),
183
+ (6, 256, 1536, 8960, 'int8', 'int8'): (256, 256, 8960),
184
+ (6, 256, 1792, 5120, 'int8', 'int8'): (256, 896, 5120),
185
+ (6, 256, 17920, 1536, 'int8', 'int8'): (256, 3584, 1536),
186
+ (6, 256, 2048, 1536, 'int8', 'int8'): (256, 2048, 1536),
187
+ (6, 256, 28672, 4096, 'int8', 'int8'): (256, 1024, 4096),
188
+ (6, 256, 3584, 18944, 'int8', 'int8'): (256, 256, 18944),
189
+ (6, 256, 3584, 3584, 'int8', 'int8'): (256, 512, 3584),
190
+ (6, 256, 3584, 8192, 'int8', 'int8'): (256, 3584, 1024),
191
+ (6, 256, 37888, 3584, 'int8', 'int8'): (256, 9472, 512),
192
+ (6, 256, 4096, 14336, 'int8', 'int8'): (256, 2048, 2048),
193
+ (6, 256, 4096, 4096, 'int8', 'int8'): (256, 512, 4096),
194
+ (6, 256, 4608, 3584, 'int8', 'int8'): (256, 512, 3584),
195
+ (6, 256, 5120, 1280, 'int8', 'int8'): (256, 2560, 1280),
196
+ (6, 256, 5120, 3456, 'int8', 'int8'): (256, 640, 3456),
197
+ (6, 256, 5120, 640, 'int8', 'int8'): (256, 2560, 640),
198
+ (6, 256, 5120, 6912, 'int8', 'int8'): (256, 640, 6912),
199
+ (6, 256, 6144, 4096, 'int8', 'int8'): (256, 2048, 2048),
200
+ (6, 256, 6912, 5120, 'int8', 'int8'): (256, 768, 5120),
201
+ (6, 256, 7168, 8192, 'int8', 'int8'): (256, 512, 8192),
202
+ (6, 256, 8192, 1024, 'int8', 'int8'): (256, 2048, 1024),
203
+ (6, 256, 8192, 3584, 'int8', 'int8'): (256, 8192, 512),
204
+ (6, 256, 896, 5120, 'int8', 'int8'): (256, 896, 2560),
205
+ (6, 32, 1024, 4096, 'int8', 'int8'): (32, 512, 4096),
206
+ (6, 32, 1024, 8192, 'int8', 'int8'): (32, 256, 8192),
207
+ (6, 32, 128, 8192, 'int8', 'int8'): (32, 128, 4096),
208
+ (6, 32, 1280, 8192, 'int8', 'int8'): (32, 1280, 2048),
209
+ (6, 32, 13824, 5120, 'int8', 'int8'): (32, 512, 5120),
210
+ (6, 32, 14336, 4096, 'int8', 'int8'): (32, 896, 4096),
211
+ (6, 32, 1536, 1536, 'int8', 'int8'): (32, 768, 1536),
212
+ (6, 32, 1536, 8960, 'int8', 'int8'): (32, 1536, 1792),
213
+ (6, 32, 1792, 5120, 'int8', 'int8'): (32, 1792, 2560),
214
+ (6, 32, 17920, 1536, 'int8', 'int8'): (32, 1792, 1536),
215
+ (6, 32, 2048, 1536, 'int8', 'int8'): (32, 1024, 1536),
216
+ (6, 32, 28672, 4096, 'int8', 'int8'): (32, 896, 4096),
217
+ (6, 32, 3584, 18944, 'int8', 'int8'): (32, 256, 18944),
218
+ (6, 32, 3584, 3584, 'int8', 'int8'): (32, 896, 3584),
219
+ (6, 32, 3584, 8192, 'int8', 'int8'): (32, 512, 8192),
220
+ (6, 32, 37888, 3584, 'int8', 'int8'): (32, 37888, 128),
221
+ (6, 32, 4096, 14336, 'int8', 'int8'): (32, 256, 14336),
222
+ (6, 32, 4096, 4096, 'int8', 'int8'): (32, 1024, 4096),
223
+ (6, 32, 4608, 3584, 'int8', 'int8'): (32, 1536, 1792),
224
+ (6, 32, 5120, 1280, 'int8', 'int8'): (32, 2560, 1280),
225
+ (6, 32, 5120, 3456, 'int8', 'int8'): (32, 1024, 3456),
226
+ (6, 32, 5120, 640, 'int8', 'int8'): (32, 2560, 640),
227
+ (6, 32, 5120, 6912, 'int8', 'int8'): (32, 512, 6912),
228
+ (6, 32, 6144, 4096, 'int8', 'int8'): (32, 768, 4096),
229
+ (6, 32, 6912, 5120, 'int8', 'int8'): (32, 768, 5120),
230
+ (6, 32, 7168, 8192, 'int8', 'int8'): (32, 512, 8192),
231
+ (6, 32, 8192, 1024, 'int8', 'int8'): (32, 8192, 256),
232
+ (6, 32, 8192, 3584, 'int8', 'int8'): (32, 4096, 896),
233
+ (6, 32, 896, 5120, 'int8', 'int8'): (32, 896, 2560),
234
+ (6, 4096, 1024, 4096, 'int8', 'int8'): (2048, 512, 4096),
235
+ (6, 4096, 1024, 8192, 'int8', 'int8'): (4096, 1024, 1024),
236
+ (6, 4096, 128, 8192, 'int8', 'int8'): (2048, 128, 8192),
237
+ (6, 4096, 1280, 8192, 'int8', 'int8'): (1024, 1280, 8192),
238
+ (6, 4096, 13824, 5120, 'int8', 'int8'): (2048, 768, 5120),
239
+ (6, 4096, 14336, 4096, 'int8', 'int8'): (1024, 2048, 4096),
240
+ (6, 4096, 1536, 1536, 'int8', 'int8'): (512, 1536, 1536),
241
+ (6, 4096, 1536, 8960, 'int8', 'int8'): (1024, 1536, 4480),
242
+ (6, 4096, 1792, 5120, 'int8', 'int8'): (2048, 1792, 5120),
243
+ (6, 4096, 17920, 1536, 'int8', 'int8'): (256, 17920, 1536),
244
+ (6, 4096, 2048, 1536, 'int8', 'int8'): (512, 2048, 1536),
245
+ (6, 4096, 28672, 4096, 'int8', 'int8'): (2048, 1792, 4096),
246
+ (6, 4096, 3584, 18944, 'int8', 'int8'): (4096, 1792, 512),
247
+ (6, 4096, 3584, 3584, 'int8', 'int8'): (4096, 512, 3584),
248
+ (6, 4096, 3584, 8192, 'int8', 'int8'): (256, 3584, 8192),
249
+ (6, 4096, 37888, 3584, 'int8', 'int8'): (2048, 1024, 3584),
250
+ (6, 4096, 4096, 14336, 'int8', 'int8'): (1024, 4096, 1792),
251
+ (6, 4096, 4096, 4096, 'int8', 'int8'): (4096, 512, 4096),
252
+ (6, 4096, 4608, 3584, 'int8', 'int8'): (4096, 512, 3584),
253
+ (6, 4096, 5120, 1280, 'int8', 'int8'): (256, 5120, 1280),
254
+ (6, 4096, 5120, 3456, 'int8', 'int8'): (4096, 512, 3456),
255
+ (6, 4096, 5120, 640, 'int8', 'int8'): (512, 5120, 640),
256
+ (6, 4096, 5120, 6912, 'int8', 'int8'): (512, 5120, 6912),
257
+ (6, 4096, 6144, 4096, 'int8', 'int8'): (4096, 512, 4096),
258
+ (6, 4096, 6912, 5120, 'int8', 'int8'): (512, 6912, 5120),
259
+ (6, 4096, 7168, 8192, 'int8', 'int8'): (1024, 1792, 8192),
260
+ (6, 4096, 8192, 1024, 'int8', 'int8'): (4096, 512, 1024),
261
+ (6, 4096, 8192, 3584, 'int8', 'int8'): (4096, 512, 3584),
262
+ (6, 4096, 896, 5120, 'int8', 'int8'): (512, 896, 5120),
263
+ (6, 512, 1024, 4096, 'int8', 'int8'): (512, 512, 2048),
264
+ (6, 512, 1024, 8192, 'int8', 'int8'): (512, 512, 8192),
265
+ (6, 512, 128, 8192, 'int8', 'int8'): (256, 128, 8192),
266
+ (6, 512, 1280, 8192, 'int8', 'int8'): (512, 256, 8192),
267
+ (6, 512, 13824, 5120, 'int8', 'int8'): (512, 768, 5120),
268
+ (6, 512, 14336, 4096, 'int8', 'int8'): (512, 896, 4096),
269
+ (6, 512, 1536, 1536, 'int8', 'int8'): (512, 768, 1536),
270
+ (6, 512, 1536, 8960, 'int8', 'int8'): (512, 256, 8960),
271
+ (6, 512, 1792, 5120, 'int8', 'int8'): (512, 1792, 1280),
272
+ (6, 512, 17920, 1536, 'int8', 'int8'): (512, 1280, 1536),
273
+ (6, 512, 2048, 1536, 'int8', 'int8'): (256, 2048, 1536),
274
+ (6, 512, 28672, 4096, 'int8', 'int8'): (512, 1024, 4096),
275
+ (6, 512, 3584, 18944, 'int8', 'int8'): (512, 256, 18944),
276
+ (6, 512, 3584, 3584, 'int8', 'int8'): (512, 1792, 3584),
277
+ (6, 512, 3584, 8192, 'int8', 'int8'): (512, 3584, 2048),
278
+ (6, 512, 37888, 3584, 'int8', 'int8'): (512, 1024, 3584),
279
+ (6, 512, 4096, 14336, 'int8', 'int8'): (512, 512, 14336),
280
+ (6, 512, 4096, 4096, 'int8', 'int8'): (512, 1024, 4096),
281
+ (6, 512, 4608, 3584, 'int8', 'int8'): (512, 768, 3584),
282
+ (6, 512, 5120, 1280, 'int8', 'int8'): (256, 5120, 1280),
283
+ (6, 512, 5120, 3456, 'int8', 'int8'): (512, 1280, 3456),
284
+ (6, 512, 5120, 640, 'int8', 'int8'): (512, 2560, 640),
285
+ (6, 512, 5120, 6912, 'int8', 'int8'): (512, 512, 6912),
286
+ (6, 512, 6144, 4096, 'int8', 'int8'): (512, 768, 4096),
287
+ (6, 512, 6912, 5120, 'int8', 'int8'): (512, 768, 5120),
288
+ (6, 512, 7168, 8192, 'int8', 'int8'): (512, 1792, 4096),
289
+ (6, 512, 8192, 1024, 'int8', 'int8'): (512, 2048, 1024),
290
+ (6, 512, 8192, 3584, 'int8', 'int8'): (512, 1024, 3584),
291
+ (6, 512, 896, 5120, 'int8', 'int8'): (512, 896, 2560),
292
+ (6, 64, 1024, 4096, 'int8', 'int8'): (64, 1024, 1024),
293
+ (6, 64, 1024, 8192, 'int8', 'int8'): (64, 1024, 2048),
294
+ (6, 64, 128, 8192, 'int8', 'int8'): (64, 128, 4096),
295
+ (6, 64, 1280, 8192, 'int8', 'int8'): (64, 1280, 2048),
296
+ (6, 64, 13824, 5120, 'int8', 'int8'): (64, 13824, 256),
297
+ (6, 64, 14336, 4096, 'int8', 'int8'): (64, 7168, 512),
298
+ (6, 64, 1536, 1536, 'int8', 'int8'): (64, 256, 1536),
299
+ (6, 64, 1536, 8960, 'int8', 'int8'): (64, 256, 8960),
300
+ (6, 64, 1792, 5120, 'int8', 'int8'): (64, 896, 5120),
301
+ (6, 64, 17920, 1536, 'int8', 'int8'): (64, 1792, 1536),
302
+ (6, 64, 2048, 1536, 'int8', 'int8'): (64, 2048, 768),
303
+ (6, 64, 28672, 4096, 'int8', 'int8'): (64, 1024, 4096),
304
+ (6, 64, 3584, 18944, 'int8', 'int8'): (64, 256, 18944),
305
+ (6, 64, 3584, 3584, 'int8', 'int8'): (64, 896, 3584),
306
+ (6, 64, 3584, 8192, 'int8', 'int8'): (64, 512, 8192),
307
+ (6, 64, 37888, 3584, 'int8', 'int8'): (64, 18944, 256),
308
+ (6, 64, 4096, 14336, 'int8', 'int8'): (64, 256, 14336),
309
+ (6, 64, 4096, 4096, 'int8', 'int8'): (64, 4096, 1024),
310
+ (6, 64, 4608, 3584, 'int8', 'int8'): (64, 768, 3584),
311
+ (6, 64, 5120, 1280, 'int8', 'int8'): (64, 2560, 1280),
312
+ (6, 64, 5120, 3456, 'int8', 'int8'): (64, 1024, 3456),
313
+ (6, 64, 5120, 640, 'int8', 'int8'): (64, 2560, 640),
314
+ (6, 64, 5120, 6912, 'int8', 'int8'): (64, 512, 6912),
315
+ (6, 64, 6144, 4096, 'int8', 'int8'): (64, 1536, 4096),
316
+ (6, 64, 6912, 5120, 'int8', 'int8'): (64, 768, 5120),
317
+ (6, 64, 7168, 8192, 'int8', 'int8'): (64, 1024, 8192),
318
+ (6, 64, 8192, 1024, 'int8', 'int8'): (64, 2048, 1024),
319
+ (6, 64, 8192, 3584, 'int8', 'int8'): (64, 1024, 3584),
320
+ (6, 64, 896, 5120, 'int8', 'int8'): (64, 896, 2560),
321
+ (6, 8192, 1024, 4096, 'int8', 'int8'): (2048, 1024, 4096),
322
+ (6, 8192, 1024, 8192, 'int8', 'int8'): (512, 1024, 8192),
323
+ (6, 8192, 128, 8192, 'int8', 'int8'): (256, 128, 8192),
324
+ (6, 8192, 1280, 8192, 'int8', 'int8'): (512, 1280, 8192),
325
+ (6, 8192, 13824, 5120, 'int8', 'int8'): (2048, 1536, 5120),
326
+ (6, 8192, 14336, 4096, 'int8', 'int8'): (2048, 1792, 4096),
327
+ (6, 8192, 1536, 1536, 'int8', 'int8'): (1024, 1536, 1536),
328
+ (6, 8192, 1536, 8960, 'int8', 'int8'): (256, 1536, 8960),
329
+ (6, 8192, 1792, 5120, 'int8', 'int8'): (1024, 1792, 2560),
330
+ (6, 8192, 17920, 1536, 'int8', 'int8'): (512, 17920, 1536),
331
+ (6, 8192, 2048, 1536, 'int8', 'int8'): (512, 2048, 1536),
332
+ (6, 8192, 28672, 4096, 'int8', 'int8'): (2048, 1792, 4096),
333
+ (6, 8192, 3584, 18944, 'int8', 'int8'): (4096, 1792, 512),
334
+ (6, 8192, 3584, 3584, 'int8', 'int8'): (2048, 1792, 3584),
335
+ (6, 8192, 3584, 8192, 'int8', 'int8'): (256, 3584, 8192),
336
+ (6, 8192, 37888, 3584, 'int8', 'int8'): (4096, 512, 3584),
337
+ (6, 8192, 4096, 14336, 'int8', 'int8'): (1024, 2048, 3584),
338
+ (6, 8192, 4096, 4096, 'int8', 'int8'): (1024, 2048, 4096),
339
+ (6, 8192, 4608, 3584, 'int8', 'int8'): (2048, 1536, 3584),
340
+ (6, 8192, 5120, 1280, 'int8', 'int8'): (2048, 1280, 1280),
341
+ (6, 8192, 5120, 3456, 'int8', 'int8'): (1024, 5120, 3456),
342
+ (6, 8192, 5120, 640, 'int8', 'int8'): (512, 5120, 640),
343
+ (6, 8192, 5120, 6912, 'int8', 'int8'): (512, 5120, 6912),
344
+ (6, 8192, 6144, 4096, 'int8', 'int8'): (1024, 6144, 4096),
345
+ (6, 8192, 6912, 5120, 'int8', 'int8'): (512, 6912, 5120),
346
+ (6, 8192, 7168, 8192, 'int8', 'int8'): (256, 1792, 8192),
347
+ (6, 8192, 8192, 1024, 'int8', 'int8'): (512, 8192, 1024),
348
+ (6, 8192, 8192, 3584, 'int8', 'int8'): (1024, 4096, 3584),
349
+ (6, 8192, 896, 5120, 'int8', 'int8'): (2048, 896, 5120),
350
+ (7, 1024, 14336, 4096, 'float8_e4m3fn', 'float8_e4m3fn'):
351
+ (1024, 1792, 4096),
352
+ (7, 1024, 25600, 5120, 'float8_e4m3fn', 'float8_e4m3fn'):
353
+ (1024, 1280, 5120),
354
+ (7, 1024, 28672, 8192, 'float8_e4m3fn', 'float8_e4m3fn'):
355
+ (512, 1024, 8192),
356
+ (7, 1024, 3072, 4096, 'float8_e4m3fn', 'float8_e4m3fn'):
357
+ (1024, 3072, 1024),
358
+ (7, 1024, 4096, 2048, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 1024,
359
+ 2048),
360
+ (7, 1024, 4096, 7168, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 4096,
361
+ 1024),
362
+ (7, 1024, 5120, 12800, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 2560,
363
+ 2560),
364
+ (7, 1024, 5120, 25600, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 2560,
365
+ 2560),
366
+ (7, 1024, 5120, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 1280,
367
+ 4096),
368
+ (7, 1024, 5120, 5120, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 1280,
369
+ 5120),
370
+ (7, 1024, 5120, 8192, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 5120, 2048),
371
+ (7, 1024, 8192, 14336, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 2048,
372
+ 2048),
373
+ (7, 1024, 8192, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 1024,
374
+ 4096),
375
+ (7, 128, 14336, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (128, 3584, 4096),
376
+ (7, 128, 25600, 5120, 'float8_e4m3fn', 'float8_e4m3fn'): (128, 3200, 5120),
377
+ (7, 128, 28672, 8192, 'float8_e4m3fn', 'float8_e4m3fn'): (128, 1792, 8192),
378
+ (7, 128, 3072, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (128, 1536, 4096),
379
+ (7, 128, 4096, 2048, 'float8_e4m3fn', 'float8_e4m3fn'): (128, 1024, 2048),
380
+ (7, 128, 4096, 7168, 'float8_e4m3fn', 'float8_e4m3fn'): (128, 1024, 7168),
381
+ (7, 128, 5120, 12800, 'float8_e4m3fn', 'float8_e4m3fn'): (128, 1024,
382
+ 12800),
383
+ (7, 128, 5120, 25600, 'float8_e4m3fn', 'float8_e4m3fn'): (128, 512, 25600),
384
+ (7, 128, 5120, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (128, 2560, 4096),
385
+ (7, 128, 5120, 5120, 'float8_e4m3fn', 'float8_e4m3fn'): (128, 1280, 5120),
386
+ (7, 128, 5120, 8192, 'float8_e4m3fn', 'float8_e4m3fn'): (128, 5120, 2048),
387
+ (7, 128, 8192, 14336, 'float8_e4m3fn', 'float8_e4m3fn'): (128, 1024,
388
+ 14336),
389
+ (7, 128, 8192, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (128, 4096, 4096),
390
+ (7, 16, 14336, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (16, 3584, 4096),
391
+ (7, 16, 25600, 5120, 'float8_e4m3fn', 'float8_e4m3fn'): (16, 3200, 5120),
392
+ (7, 16, 28672, 8192, 'float8_e4m3fn', 'float8_e4m3fn'): (16, 2048, 8192),
393
+ (7, 16, 3072, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (16, 1536, 4096),
394
+ (7, 16, 4096, 2048, 'float8_e4m3fn', 'float8_e4m3fn'): (16, 4096, 1024),
395
+ (7, 16, 4096, 7168, 'float8_e4m3fn', 'float8_e4m3fn'): (16, 4096, 1792),
396
+ (7, 16, 5120, 12800, 'float8_e4m3fn', 'float8_e4m3fn'): (16, 1024, 12800),
397
+ (7, 16, 5120, 25600, 'float8_e4m3fn', 'float8_e4m3fn'): (16, 640, 25600),
398
+ (7, 16, 5120, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (16, 1280, 4096),
399
+ (7, 16, 5120, 5120, 'float8_e4m3fn', 'float8_e4m3fn'): (16, 1280, 5120),
400
+ (7, 16, 5120, 8192, 'float8_e4m3fn', 'float8_e4m3fn'): (16, 1280, 8192),
401
+ (7, 16, 8192, 14336, 'float8_e4m3fn', 'float8_e4m3fn'): (16, 8192, 1024),
402
+ (7, 16, 8192, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (16, 2048, 4096),
403
+ (7, 2048, 14336, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 2048,
404
+ 4096),
405
+ (7, 2048, 25600, 5120, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 2560,
406
+ 5120),
407
+ (7, 2048, 28672, 8192, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 4096,
408
+ 1024),
409
+ (7, 2048, 3072, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 1536,
410
+ 4096),
411
+ (7, 2048, 4096, 2048, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 2048,
412
+ 2048),
413
+ (7, 2048, 4096, 7168, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 4096,
414
+ 1024),
415
+ (7, 2048, 5120, 12800, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 2560,
416
+ 2560),
417
+ (7, 2048, 5120, 25600, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 2560,
418
+ 2560),
419
+ (7, 2048, 5120, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 1280,
420
+ 4096),
421
+ (7, 2048, 5120, 5120, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 2560,
422
+ 2560),
423
+ (7, 2048, 5120, 8192, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 2560,
424
+ 2048),
425
+ (7, 2048, 8192, 14336, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 4096,
426
+ 1024),
427
+ (7, 2048, 8192, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 2048,
428
+ 4096),
429
+ (7, 256, 14336, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (256, 2048, 4096),
430
+ (7, 256, 25600, 5120, 'float8_e4m3fn', 'float8_e4m3fn'): (256, 2560, 5120),
431
+ (7, 256, 28672, 8192, 'float8_e4m3fn', 'float8_e4m3fn'): (256, 7168, 2048),
432
+ (7, 256, 3072, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (256, 3072, 2048),
433
+ (7, 256, 4096, 2048, 'float8_e4m3fn', 'float8_e4m3fn'): (256, 2048, 2048),
434
+ (7, 256, 4096, 7168, 'float8_e4m3fn', 'float8_e4m3fn'): (256, 1024, 7168),
435
+ (7, 256, 5120, 12800, 'float8_e4m3fn', 'float8_e4m3fn'): (256, 5120, 2560),
436
+ (7, 256, 5120, 25600, 'float8_e4m3fn', 'float8_e4m3fn'): (256, 5120, 2560),
437
+ (7, 256, 5120, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (256, 1280, 4096),
438
+ (7, 256, 5120, 5120, 'float8_e4m3fn', 'float8_e4m3fn'): (256, 1280, 5120),
439
+ (7, 256, 5120, 8192, 'float8_e4m3fn', 'float8_e4m3fn'): (256, 1280, 8192),
440
+ (7, 256, 8192, 14336, 'float8_e4m3fn', 'float8_e4m3fn'): (256, 8192, 1792),
441
+ (7, 256, 8192, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (256, 2048, 4096),
442
+ (7, 32, 14336, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (32, 1792, 4096),
443
+ (7, 32, 25600, 5120, 'float8_e4m3fn', 'float8_e4m3fn'): (32, 2560, 5120),
444
+ (7, 32, 28672, 8192, 'float8_e4m3fn', 'float8_e4m3fn'): (32, 2048, 8192),
445
+ (7, 32, 3072, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (32, 3072, 2048),
446
+ (7, 32, 4096, 2048, 'float8_e4m3fn', 'float8_e4m3fn'): (32, 1024, 2048),
447
+ (7, 32, 4096, 7168, 'float8_e4m3fn', 'float8_e4m3fn'): (32, 2048, 7168),
448
+ (7, 32, 5120, 12800, 'float8_e4m3fn', 'float8_e4m3fn'): (32, 1280, 12800),
449
+ (7, 32, 5120, 25600, 'float8_e4m3fn', 'float8_e4m3fn'): (32, 640, 25600),
450
+ (7, 32, 5120, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (32, 1280, 4096),
451
+ (7, 32, 5120, 5120, 'float8_e4m3fn', 'float8_e4m3fn'): (32, 1280, 5120),
452
+ (7, 32, 5120, 8192, 'float8_e4m3fn', 'float8_e4m3fn'): (32, 2560, 4096),
453
+ (7, 32, 8192, 14336, 'float8_e4m3fn', 'float8_e4m3fn'): (32, 2048, 7168),
454
+ (7, 32, 8192, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (32, 4096, 4096),
455
+ (7, 4096, 14336, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 3584,
456
+ 4096),
457
+ (7, 4096, 25600, 5120, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 1280,
458
+ 5120),
459
+ (7, 4096, 28672, 8192, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 1024,
460
+ 8192),
461
+ (7, 4096, 3072, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 3072, 4096),
462
+ (7, 4096, 4096, 2048, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 4096, 2048),
463
+ (7, 4096, 4096, 7168, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 4096,
464
+ 1024),
465
+ (7, 4096, 5120, 12800, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 2560,
466
+ 2560),
467
+ (7, 4096, 5120, 25600, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 2560,
468
+ 2560),
469
+ (7, 4096, 5120, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 1280,
470
+ 4096),
471
+ (7, 4096, 5120, 5120, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 1280,
472
+ 5120),
473
+ (7, 4096, 5120, 8192, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 2560,
474
+ 2048),
475
+ (7, 4096, 8192, 14336, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 4096,
476
+ 1024),
477
+ (7, 4096, 8192, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 1024,
478
+ 4096),
479
+ (7, 512, 14336, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 2048, 4096),
480
+ (7, 512, 25600, 5120, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 2560, 5120),
481
+ (7, 512, 28672, 8192, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 512, 8192),
482
+ (7, 512, 3072, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 3072, 1024),
483
+ (7, 512, 4096, 2048, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 2048, 2048),
484
+ (7, 512, 4096, 7168, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 1024, 7168),
485
+ (7, 512, 5120, 12800, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 512, 12800),
486
+ (7, 512, 5120, 25600, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 2560, 2560),
487
+ (7, 512, 5120, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 2560, 2048),
488
+ (7, 512, 5120, 5120, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 1280, 5120),
489
+ (7, 512, 5120, 8192, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 1024, 8192),
490
+ (7, 512, 8192, 14336, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 4096, 2048),
491
+ (7, 512, 8192, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 1024, 4096),
492
+ (7, 64, 14336, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (64, 14336, 1024),
493
+ (7, 64, 25600, 5120, 'float8_e4m3fn', 'float8_e4m3fn'): (64, 6400, 2560),
494
+ (7, 64, 28672, 8192, 'float8_e4m3fn', 'float8_e4m3fn'): (64, 1792, 8192),
495
+ (7, 64, 3072, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (64, 1536, 4096),
496
+ (7, 64, 4096, 2048, 'float8_e4m3fn', 'float8_e4m3fn'): (64, 2048, 2048),
497
+ (7, 64, 4096, 7168, 'float8_e4m3fn', 'float8_e4m3fn'): (64, 1024, 7168),
498
+ (7, 64, 5120, 12800, 'float8_e4m3fn', 'float8_e4m3fn'): (64, 1024, 12800),
499
+ (7, 64, 5120, 25600, 'float8_e4m3fn', 'float8_e4m3fn'): (64, 640, 25600),
500
+ (7, 64, 5120, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (64, 2560, 4096),
501
+ (7, 64, 5120, 5120, 'float8_e4m3fn', 'float8_e4m3fn'): (64, 1280, 5120),
502
+ (7, 64, 5120, 8192, 'float8_e4m3fn', 'float8_e4m3fn'): (64, 1280, 8192),
503
+ (7, 64, 8192, 14336, 'float8_e4m3fn', 'float8_e4m3fn'): (64, 1024, 14336),
504
+ (7, 64, 8192, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (64, 2048, 4096),
505
+ (7, 8192, 14336, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 1024,
506
+ 4096),
507
+ (7, 8192, 25600, 5120, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 1280,
508
+ 5120),
509
+ (7, 8192, 28672, 8192, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 2048,
510
+ 2048),
511
+ (7, 8192, 3072, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 1536,
512
+ 4096),
513
+ (7, 8192, 4096, 2048, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 4096, 2048),
514
+ (7, 8192, 4096, 7168, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 4096,
515
+ 1024),
516
+ (7, 8192, 5120, 12800, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 2560,
517
+ 3200),
518
+ (7, 8192, 5120, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 1280,
519
+ 4096),
520
+ (7, 8192, 5120, 5120, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 1280,
521
+ 5120),
522
+ (7, 8192, 5120, 8192, 'float8_e4m3fn', 'float8_e4m3fn'): (512, 5120, 2048),
523
+ (7, 8192, 8192, 14336, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 4096,
524
+ 1024),
525
+ (7, 8192, 8192, 4096, 'float8_e4m3fn', 'float8_e4m3fn'): (1024, 2048,
526
+ 4096),
527
+ # go/keep-sorted end
528
+ }
529
+
530
+ TUNED_BLOCK_SIZES: dict[TunedKey, TunedValue] = {
531
+ TunedKey(*key): TunedValue(*value)
532
+ for key, value in TUNED_BLOCK_SIZES_RAW.items()
533
+ }
534
+
535
+ DEVICE_VMEM_LIMIT = {6: 96 * 1024 * 1024, 7: 48 * 1024 * 1024}
536
+
537
+
538
+ def get_device_vmem_limit() -> int:
539
+ tpu_version = get_tpu_version()
540
+ if tpu_version not in DEVICE_VMEM_LIMIT:
541
+ logger.warning_once(
542
+ 'VMEM limit for TPU version %d not found. Using default VMEM limit '
543
+ 'of 96MiB', tpu_version)
544
+ return 96 * 1024 * 1024
545
+ return DEVICE_VMEM_LIMIT[tpu_version]
546
+
547
+
548
+ def get_tpu_version() -> int:
549
+ """Returns the numeric version of the TPU, or -1 if not on TPU."""
550
+ kind = jax.devices()[0].device_kind
551
+ match = re.match(r'^TPU[^\d]*(\d+)', kind)
552
+ if match is None:
553
+ return -1
554
+ return int(match.group(1))
555
+
556
+
557
+ def get_key(
558
+ n_batch: int,
559
+ n_out: int,
560
+ n_in: int,
561
+ x_q_dtype: str,
562
+ w_q_dtype: str,
563
+ ) -> TunedKey:
564
+ """Returns the key for the given parameters."""
565
+ return TunedKey(
566
+ get_tpu_version(),
567
+ n_batch,
568
+ n_out,
569
+ n_in,
570
+ x_q_dtype,
571
+ w_q_dtype,
572
+ )
573
+
574
+
575
+ def get_tuned_block_sizes(
576
+ n_batch: int,
577
+ n_out: int,
578
+ n_in: int,
579
+ x_q_dtype: str,
580
+ w_q_dtype: str,
581
+ ) -> TunedValue:
582
+ """Retrieve the tuned block sizes for the given parameters.
583
+
584
+ Args:
585
+ n_batch: The batch size.
586
+ n_out: The number of output features.
587
+ n_in: The number of input features.
588
+ x_q_dtype: The data type of the activation ('int8' or 'float8_e4m3fn').
589
+ w_q_dtype: The data type of the weight ('int8' or 'float8_e4m3fn').
590
+
591
+ Returns:
592
+ tuple: A tuple containing the batch_block_size, out_block_size, and
593
+ in_block_size.
594
+ """
595
+ key = get_key(
596
+ n_batch,
597
+ n_out,
598
+ n_in,
599
+ x_q_dtype,
600
+ w_q_dtype,
601
+ )
602
+ tuned_value = TUNED_BLOCK_SIZES.get(key)
603
+ if tuned_value is None:
604
+ logger.warning_once(
605
+ 'Couldn`t find tuned sizes for the quantized matmul kernel with %s',
606
+ key)
607
+ return TunedValue(128, 128, 128)
608
+ else:
609
+ return tuned_value