tpu-inference 0.11.1.dev202511130813__py3-none-any.whl → 0.11.1.dev202511180814__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.

Files changed (37) hide show
  1. tests/test_envs.py +182 -0
  2. tests/test_utils.py +23 -14
  3. tpu_inference/core/core_tpu.py +17 -9
  4. tpu_inference/executors/ray_distributed_executor.py +24 -11
  5. tpu_inference/kernels/ragged_paged_attention/v3/kernel_hd64.py +33 -10
  6. tpu_inference/kernels/ragged_paged_attention/v3/tuned_block_sizes.py +7 -0
  7. tpu_inference/layers/{jax → common}/attention_interface.py +1 -1
  8. tpu_inference/layers/common/quant_methods.py +8 -0
  9. tpu_inference/layers/jax/attention/attention.py +1 -1
  10. tpu_inference/layers/jax/sample/rejection_sampler.py +1 -1
  11. tpu_inference/layers/jax/sample/sampling.py +2 -2
  12. tpu_inference/layers/vllm/attention.py +1 -1
  13. tpu_inference/layers/vllm/quantization/__init__.py +7 -3
  14. tpu_inference/layers/vllm/quantization/awq.py +4 -3
  15. tpu_inference/layers/vllm/quantization/compressed_tensors/compressed_tensors.py +4 -2
  16. tpu_inference/layers/vllm/quantization/mxfp4.py +266 -0
  17. tpu_inference/layers/vllm/quantization/unquantized.py +4 -3
  18. tpu_inference/models/common/model_loader.py +3 -2
  19. tpu_inference/models/jax/llama3.py +2 -2
  20. tpu_inference/models/jax/phi3.py +1 -1
  21. tpu_inference/models/jax/qwen2.py +1 -1
  22. tpu_inference/models/jax/qwen2_5_vl.py +2 -2
  23. tpu_inference/models/jax/qwen3.py +1 -1
  24. tpu_inference/models/vllm/vllm_model_wrapper.py +22 -10
  25. tpu_inference/platforms/tpu_platform.py +12 -5
  26. tpu_inference/runner/compilation_manager.py +4 -2
  27. tpu_inference/runner/kv_cache.py +1 -1
  28. tpu_inference/runner/tpu_runner.py +31 -7
  29. tpu_inference/utils.py +2 -2
  30. tpu_inference/worker/tpu_worker.py +1 -1
  31. {tpu_inference-0.11.1.dev202511130813.dist-info → tpu_inference-0.11.1.dev202511180814.dist-info}/METADATA +1 -1
  32. {tpu_inference-0.11.1.dev202511130813.dist-info → tpu_inference-0.11.1.dev202511180814.dist-info}/RECORD +37 -34
  33. /tpu_inference/layers/{jax → common}/binary_search.py +0 -0
  34. /tpu_inference/layers/{jax → common}/sharding.py +0 -0
  35. {tpu_inference-0.11.1.dev202511130813.dist-info → tpu_inference-0.11.1.dev202511180814.dist-info}/WHEEL +0 -0
  36. {tpu_inference-0.11.1.dev202511130813.dist-info → tpu_inference-0.11.1.dev202511180814.dist-info}/licenses/LICENSE +0 -0
  37. {tpu_inference-0.11.1.dev202511130813.dist-info → tpu_inference-0.11.1.dev202511180814.dist-info}/top_level.txt +0 -0
{tpu_inference-0.11.1.dev202511130813.dist-info → tpu_inference-0.11.1.dev202511180814.dist-info}/METADATA RENAMED
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: tpu_inference
3
- Version: 0.11.1.dev202511130813
3
+ Version: 0.11.1.dev202511180814
4
4
  Author: tpu_inference Contributors
5
5
  Classifier: Development Status :: 3 - Alpha
6
6
  Classifier: Intended Audience :: Developers
{tpu_inference-0.11.1.dev202511130813.dist-info → tpu_inference-0.11.1.dev202511180814.dist-info}/RECORD RENAMED
@@ -1,8 +1,9 @@
1
1
  tests/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
2
2
  tests/test_base.py,sha256=Ct5WFRMHL7IHEIxk8FrzAvO8m0xFuDpzDBKkAKKAL2Q,7341
3
+ tests/test_envs.py,sha256=Woyfp_d5HS-uTGo4_u9dYlBbgmhfIEoFb-Rx_k7YXD4,6298
3
4
  tests/test_quantization.py,sha256=IT5ASyS1uuWcxc22kRtBcA-V4j3Z3hb7pMztm3GOlBs,34445
4
5
  tests/test_tpu_info.py,sha256=ZrwlMsp8ffITkS_b8Q1t_QG-a-WVAd4NUcjHhGibcsI,4670
5
- tests/test_utils.py,sha256=szRg4UB36RcgIvbEd9xMhKYbWi-O4XAUWGJlIU6FJ9E,7983
6
+ tests/test_utils.py,sha256=Mta5ZzYCgRAh1-BjcOvvx9iQ9DnnXLps7oDHxVQp2yE,8236
6
7
  tests/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
7
8
  tests/core/test_core_tpu.py,sha256=r496rk1eOsK_F4nvm9zprl_T-RcO6eCUb7LuVReOZno,21413
8
9
  tests/core/test_disagg_executor.py,sha256=QdE2YZs08EyDDCmSjhiXkXqQ9BJTgO6csr_E1xkkfSg,2256
@@ -28,9 +29,9 @@ tpu_inference/env_override.py,sha256=pmL7lfs_rGCP92ya3wuWuudsCYeOMZ6tFZY82A4KkQc
28
29
  tpu_inference/envs.py,sha256=MTT_Pdtd6cAcciYjv1OekEmvspaq3SYL0oR_jDkQ_aE,3948
29
30
  tpu_inference/logger.py,sha256=HQCz7NefmbturuhOC7-3Ixbtcdgoz4g9FHh2RB6o8cc,334
30
31
  tpu_inference/tpu_info.py,sha256=9UohshkndR6dZpGWpWXfTD4qvIVdVgHf0yOoSEkLTrw,2276
31
- tpu_inference/utils.py,sha256=LWEshJgUdB20H2fDA-QI-Sk4EP7PD_FWvW3Mrqb-k8M,10054
32
+ tpu_inference/utils.py,sha256=iGPY147jP_8AKMu3g7vYTndjJJiOrK_4opA0JWtws5Q,10068
32
33
  tpu_inference/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
33
- tpu_inference/core/core_tpu.py,sha256=JdN4-xaxSWnzY4T181SCnbZ5HEnwQ5IifYA9ybF4pWo,32710
34
+ tpu_inference/core/core_tpu.py,sha256=WDD3koE_j1QhWS2BbMA2aQOZayPZm4tYPvzL4YCX2jY,33294
34
35
  tpu_inference/core/disagg_executor.py,sha256=HZpgYMVxRxm0RQxO4l8IDYBWJ6Z3Tac6xavc5otcirc,4657
35
36
  tpu_inference/core/disagg_utils.py,sha256=ufWNFWQ5n4YnZpPOtoReHlYo4dlN7AbIqCyqS4an0t4,1572
36
37
  tpu_inference/core/sched/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -40,7 +41,7 @@ tpu_inference/distributed/jax_parallel_state.py,sha256=5_xCwcL03lFPUoSO_OP7hIVKp
40
41
  tpu_inference/distributed/tpu_connector.py,sha256=Zah46Sm5iOuh72SzXw69NxMc0MLnqsLEpe2BfDhpnqA,29731
41
42
  tpu_inference/distributed/utils.py,sha256=RwFQi8G4TzN1g9RjQu0pb5JxSc_jhoIZVsFJo0uHjxo,1513
42
43
  tpu_inference/executors/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
43
- tpu_inference/executors/ray_distributed_executor.py,sha256=UgJP-XSgDPKDj_mkVQ16XrRN96juVpnFl6fdWEyFL_Y,15249
44
+ tpu_inference/executors/ray_distributed_executor.py,sha256=ZMuVUwmroi7UUZs3u67OsOwUIkxNDz9IszUPG20F18E,15904
44
45
  tpu_inference/experimental/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
45
46
  tpu_inference/experimental/llama3_jax_stashed.py,sha256=YK1oSIfto9ALo-HB45XfSrbq9XgVbE4m2C-9zRwmSzI,10913
46
47
  tpu_inference/kernels/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -67,26 +68,27 @@ tpu_inference/kernels/ragged_paged_attention/v2/ragged_kv_cache_update.py,sha256
67
68
  tpu_inference/kernels/ragged_paged_attention/v2/tuned_block_sizes.py,sha256=mw80bXBGenroGdrITV0F_EaI2s-Z9KWwqU9WodvJg14,97919
68
69
  tpu_inference/kernels/ragged_paged_attention/v3/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
69
70
  tpu_inference/kernels/ragged_paged_attention/v3/kernel.py,sha256=tlP6121yfXaukx_RQroHlHcZnbKPyyum0lAcvT0B_Pk,56132
70
- tpu_inference/kernels/ragged_paged_attention/v3/kernel_hd64.py,sha256=DFVdIIKmyufu_4b-3YhxI56jt0O1cJ3JsVl-2DDZHv4,55350
71
- tpu_inference/kernels/ragged_paged_attention/v3/tuned_block_sizes.py,sha256=leTS75aq99N1Zuv6wB5yLdkfYnEtrBDVI4z_jOKnjL0,142012
71
+ tpu_inference/kernels/ragged_paged_attention/v3/kernel_hd64.py,sha256=pD1Pte3neoLAxE3I3-VyV_4FuqgCHeAHGzEjMVt0MMk,56004
72
+ tpu_inference/kernels/ragged_paged_attention/v3/tuned_block_sizes.py,sha256=k3LwduhZO85cJ-pSgnGN0c2Nn8eNeQq4eA94KUXJzMw,142198
72
73
  tpu_inference/kernels/ragged_paged_attention/v3/tuned_block_sizes_hd64.py,sha256=P3_ivi8iUz5QMU_3pgpl4Bkbmn0q0NpDtVJX39haRQA,11208
73
74
  tpu_inference/kernels/ragged_paged_attention/v3/util.py,sha256=1N_ozjKboDYLteFJndWoLXNudj2z53rGXMkELa5Z9tY,1102
74
75
  tpu_inference/layers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
75
76
  tpu_inference/layers/common/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
77
+ tpu_inference/layers/common/attention_interface.py,sha256=CImMS8tuWgvaRY9YbGS3pY7OBnzeJ4Jla7LRFb4Xoa4,13224
76
78
  tpu_inference/layers/common/attention_metadata.py,sha256=St8ZatbY1D7xQACKJH459jMgp3oTP3AQ36mi9FZdrPU,850
79
+ tpu_inference/layers/common/binary_search.py,sha256=ZQi-z1wG6WTcfVQXeTGOZokX4K1DSf9kCzqfrhEU8lk,12320
80
+ tpu_inference/layers/common/quant_methods.py,sha256=mQSxZ44-QQtm22C_8ViejnP1cP2Dv6yc2YaP6oMKJeQ,185
81
+ tpu_inference/layers/common/sharding.py,sha256=wBqdkXZSWfnnH8pkJtyW2DSqmAe_V4Vxi0iMPaXq0Z0,25185
77
82
  tpu_inference/layers/jax/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
78
- tpu_inference/layers/jax/attention_interface.py,sha256=1jlvSZWaP6DuPVtb1W_KPw4-Qi68BikOBNLLcpygupY,13221
79
83
  tpu_inference/layers/jax/base.py,sha256=Vhts6ZMwNCZ8LbnEXeB0rl3nHdS5hDJWX7HEa7Fl7yE,5775
80
- tpu_inference/layers/jax/binary_search.py,sha256=ZQi-z1wG6WTcfVQXeTGOZokX4K1DSf9kCzqfrhEU8lk,12320
81
84
  tpu_inference/layers/jax/constants.py,sha256=NcYg0zAf3ClfP7YMYdYu_F1GngOzZaIxIAHBZDunKw4,2755
82
85
  tpu_inference/layers/jax/layers.py,sha256=yv_lC2tbJuzVL-OaXYooX82Ys8hWZATeH9M78coJ3VI,10633
83
86
  tpu_inference/layers/jax/misc.py,sha256=znKv1Nuq_LgYpaIu0qlzUVDgQWnjjG7aqPJGM8kuwcw,566
84
87
  tpu_inference/layers/jax/rope.py,sha256=i2E7pRLWgOaFLbeo8_phZwKQWJW7ohAyl69E2V2Mc2U,11349
85
88
  tpu_inference/layers/jax/rope_interface.py,sha256=X0SruXizlCHGnssFujC1pL07UC4Vsp7-gdBy_Q7JZhI,8375
86
- tpu_inference/layers/jax/sharding.py,sha256=wBqdkXZSWfnnH8pkJtyW2DSqmAe_V4Vxi0iMPaXq0Z0,25185
87
89
  tpu_inference/layers/jax/transformer_block.py,sha256=ufv-yfVDmRP_Ynrx3UX9xj-x0PkNw_tQ-0N0eYf4i7M,3917
88
90
  tpu_inference/layers/jax/attention/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
89
- tpu_inference/layers/jax/attention/attention.py,sha256=bWXMtF8TToiRyZ3SvJGQnD2urZTbuX_omHVXpQdn0fs,10082
91
+ tpu_inference/layers/jax/attention/attention.py,sha256=DJFDkpQc9SDD156wVPFw3r2XaBgb44QNJ8OcdONaF5g,10085
90
92
  tpu_inference/layers/jax/attention/deepseek_v3_attention.py,sha256=YlagoBMwINv2KRH1dr4oEcH_cQ9QMPB55nO2FQZsWs0,14010
91
93
  tpu_inference/layers/jax/attention/gpt_oss_attention.py,sha256=rkrEv4aNZxtAGcXd1HXHUxhNeDNAd9nWTEZOKWSI8cA,8725
92
94
  tpu_inference/layers/jax/attention/llama4_attention.py,sha256=VvUmfBxQEbHf3F2BrcYDUnq5abj7CSDYeRsNx_eVAh0,6162
@@ -95,20 +97,21 @@ tpu_inference/layers/jax/moe/deepseek_v3_moe.py,sha256=Q6CuwwiZtWYm6iUee1wJoDJrw
95
97
  tpu_inference/layers/jax/moe/gpt_oss_moe.py,sha256=Rx5b1jg2XMm7Xx9hrjgvyhscaJ_zGbVMHmeEiLh7kIQ,6196
96
98
  tpu_inference/layers/jax/moe/moe.py,sha256=cA8R1rjbBwNEoNlsPWjeIBB9nvaRDwlEdwQTVg6lTpY,8762
97
99
  tpu_inference/layers/jax/sample/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
98
- tpu_inference/layers/jax/sample/rejection_sampler.py,sha256=IRfVWjkbVXp9Sv1YrGMMh-LYx1AwbY-3FTXEO1-Ue9g,20423
99
- tpu_inference/layers/jax/sample/sampling.py,sha256=dVOcMdmPdAEsupPk96tCaZecIWUiDej0DiVnwaH9ckQ,3308
100
+ tpu_inference/layers/jax/sample/rejection_sampler.py,sha256=nI5s0E73xkqDIu2hTljIXt23B1Q-gRnC1myoQpGDJrQ,20426
101
+ tpu_inference/layers/jax/sample/sampling.py,sha256=C30KgmdOVSaagvHhbfLgVJtVQmJo86CbHPa4h36Vn70,3314
100
102
  tpu_inference/layers/jax/sample/sampling_metadata.py,sha256=Gd835LNWfGM0NRQBVBqEv0nPwt5q9F4AdFym0CUS1fw,2561
101
103
  tpu_inference/layers/vllm/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
102
- tpu_inference/layers/vllm/attention.py,sha256=JxEQ8ql_97zbQzukIbfUYq50-2k81VUG1Km_YV_RUtg,7363
104
+ tpu_inference/layers/vllm/attention.py,sha256=wbJpcgqEAuIirv5PIULbiP-ggMKjmTanbB7Dg0BVYv4,7366
103
105
  tpu_inference/layers/vllm/fused_moe.py,sha256=XZt2CPUz00qZzDcyfBFz6buhVzmGL1amHalHJALl9zw,18945
104
106
  tpu_inference/layers/vllm/linear_common.py,sha256=_YlJtbdaYcck_j-gFLos_k0ycktVWxT8Qo57tR2YqJ8,7749
105
107
  tpu_inference/layers/vllm/sharding.py,sha256=WTx1tF_7R99AdyE-lL7HQJ378hAafeI-JVRsugAvwn4,9177
106
- tpu_inference/layers/vllm/quantization/__init__.py,sha256=Tz44kUZTdNFu5Dmu48aQ-9f7ioWjbUWS0eVYURXZ17E,1535
107
- tpu_inference/layers/vllm/quantization/awq.py,sha256=ar8x1CPTPvfcf4wbuBC1XVh4pjtSUchoYWnbkZKH3CQ,8412
108
+ tpu_inference/layers/vllm/quantization/__init__.py,sha256=SEppGayBzzQ5tsXLSy99aqilkAawQwYxnv2alCg6-ZU,1777
109
+ tpu_inference/layers/vllm/quantization/awq.py,sha256=-8ZmjGvSKJB6_JuwSctNWt8xHWq4VSvK_AK9iahlgCo,8495
108
110
  tpu_inference/layers/vllm/quantization/common.py,sha256=wm3pge6XMTMsLK7_SSdgBP0PvQzz-1mrqN2I6xMqzrc,4218
109
- tpu_inference/layers/vllm/quantization/unquantized.py,sha256=id6d_IZIhDIvmaH3ANtmLiy4U_uY_AYAf4KTvfs3nmc,14900
111
+ tpu_inference/layers/vllm/quantization/mxfp4.py,sha256=KwGoqIiPkd6FplGuYAKi4uX5A8MPlZqq99MVPchXyi4,11561
112
+ tpu_inference/layers/vllm/quantization/unquantized.py,sha256=Q1v1ZbSIDmaoOg97Ehv6rA5CnSf6nTP40xDBMmHHeLw,15054
110
113
  tpu_inference/layers/vllm/quantization/compressed_tensors/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
111
- tpu_inference/layers/vllm/quantization/compressed_tensors/compressed_tensors.py,sha256=uKaauZhaRDcMqd8_NyQoFs9BazMOFix3nIuutbLHHbU,5123
114
+ tpu_inference/layers/vllm/quantization/compressed_tensors/compressed_tensors.py,sha256=6idEyy3e849fZ1UeNvc9eSHYX7e6qvohrJa_d_D9MBk,5285
112
115
  tpu_inference/layers/vllm/quantization/compressed_tensors/compressed_tensors_moe.py,sha256=FM901QhyhJRC8CuMeICzCVVERvBHbhruRxYW0EQ570s,8820
113
116
  tpu_inference/layers/vllm/quantization/compressed_tensors/schemes/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
114
117
  tpu_inference/layers/vllm/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_fp8.py,sha256=6sQvsxiWdi5Vte8V9vrQ2abaqGqWpq-mtzU7lGAo-ac,8759
@@ -123,18 +126,18 @@ tpu_inference/mock/vllm_logger.py,sha256=vUGnN5nKT--ZvU15YCzODUM_FGiXKhcrrjDGjeN
123
126
  tpu_inference/mock/vllm_logging_utils.py,sha256=TEUmKj3xHiLzHBnFqAujcxH0t2hBQ04sUaho2RyORnk,486
124
127
  tpu_inference/models/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
125
128
  tpu_inference/models/common/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
126
- tpu_inference/models/common/model_loader.py,sha256=AwukmGaUq2wv3OnFHUU-nwdAnKLG_eGw7PYY5CNrNNI,18225
129
+ tpu_inference/models/common/model_loader.py,sha256=VgxM2OODb0-69dexv4aNJ4g24Nrx5sj_ra4XStkhl14,18289
127
130
  tpu_inference/models/jax/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
128
131
  tpu_inference/models/jax/deepseek_v3.py,sha256=SKOHVEC-_2NLxBnzBzbu5tu0d6FTlAEiI1EefGaO2QE,40047
129
132
  tpu_inference/models/jax/gpt_oss.py,sha256=Vw4LRB5Kp6hbA2hjZGFS8kiEqOCjf881XH2JNtu2S1I,20924
130
133
  tpu_inference/models/jax/jax_intermediate_tensor.py,sha256=Pxu1PCV5LN5X58aYVkPiohcXZIeKVim2oqvrS_cVgw4,2604
131
- tpu_inference/models/jax/llama3.py,sha256=YUG0S0Y6cy7PLcq0cpmDsGWbOZIhZzzyObRQdmUUxkg,13420
134
+ tpu_inference/models/jax/llama3.py,sha256=w99DAfipGS9HyX2ZRwqyYLxC3oa0ew5eEQ6EXlMMf18,13426
132
135
  tpu_inference/models/jax/llama4.py,sha256=wf2Sp2iYViaYD5rSfv3_ryO6gYuYM5XaOyvghaP4OCY,29631
133
136
  tpu_inference/models/jax/llama_eagle3.py,sha256=STUkAK6XEA7JM3i_Lx36-t5BhkAGeW_xYiq3zYhHP1A,12297
134
- tpu_inference/models/jax/phi3.py,sha256=Oz68PE2Z1t8wTed95_w0KMIXfnfV72ZwXugNOdWOV5w,13576
135
- tpu_inference/models/jax/qwen2.py,sha256=RYb0hMKzPnFOAyhqbztoNlSrFIlRa74fYqSNecA2VOY,13354
136
- tpu_inference/models/jax/qwen2_5_vl.py,sha256=J4-AjeS_igJdxYCjTwS0HShiEfwQUMwrHxjlWvMw0ok,43939
137
- tpu_inference/models/jax/qwen3.py,sha256=SOL-Pvp56IrMxqXpPf5EFacBI6AJNlqf4Zrr1pkabGw,10994
137
+ tpu_inference/models/jax/phi3.py,sha256=TpP3Nvr1myW_Qd8xNrLP1VmXtq7BuTcWNayJitskFd0,13579
138
+ tpu_inference/models/jax/qwen2.py,sha256=P_x_Qygf-nanmF8Uufk4c-qLNxP4RAk4yuqSF8VwbxE,13357
139
+ tpu_inference/models/jax/qwen2_5_vl.py,sha256=fvMgM5GfUn5EECaMbR0z37mmbCHphAT1AvWPvGkhVn4,43942
140
+ tpu_inference/models/jax/qwen3.py,sha256=lr3TIIQKmNgWFDFxwuPsVOypqBijkqrpnNCopVg4iBo,10997
138
141
  tpu_inference/models/jax/utils/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
139
142
  tpu_inference/models/jax/utils/file_utils.py,sha256=NOuSC3YFnZpf3CZgYdghbbiNYJt42zgjlEYbOZIVct4,2840
140
143
  tpu_inference/models/jax/utils/multi_modal_utils.py,sha256=rrIrQWidkUnGilBHKNpdYh7_2BkvnAaqanXjC81GNcg,6156
@@ -147,30 +150,30 @@ tpu_inference/models/jax/utils/quantization/configs/fp8_default.yaml,sha256=b7Sy
147
150
  tpu_inference/models/jax/utils/quantization/configs/int8_all_modules_w_only.yaml,sha256=0Qwij71zj9k6rmrUNd8Q5df9YYfkoJ1ZkgMAHxQy81k,128
148
151
  tpu_inference/models/jax/utils/quantization/configs/int8_default.yaml,sha256=lGec0UwwxmNPNgKPSsTsCMSXNJjhw507KMtM2NsSCMw,152
149
152
  tpu_inference/models/vllm/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
150
- tpu_inference/models/vllm/vllm_model_wrapper.py,sha256=ERxj-cm-pmYpT9eiL-E3OxeaQDEDrH_Vs0iUS9nCU9s,11424
153
+ tpu_inference/models/vllm/vllm_model_wrapper.py,sha256=o3oJ7Uhu-vSJEFHHifF8e0Q7dULRKJ2GRsT1qAN6PWY,12099
151
154
  tpu_inference/models/vllm/vllm_model_wrapper_context.py,sha256=yxlJHPmRQIAwlb1MmHK3xfXokgIkJ-evNU4PgyoJUdg,1187
152
155
  tpu_inference/platforms/__init__.py,sha256=lQCrKddS_GcGpCbeogvz9zOZD1mQw5bBsiw8On46qFQ,74
153
- tpu_inference/platforms/tpu_platform.py,sha256=bdo_zlRqrhccpaz6zOdH18cU8kq6tGKgR1xJJehsVrc,10131
156
+ tpu_inference/platforms/tpu_platform.py,sha256=AYFr1Q7VUN76wcdgOe_wZuVIHgp2U8isBJ3iHrYqt0M,10530
154
157
  tpu_inference/runner/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
155
158
  tpu_inference/runner/block_table.py,sha256=K3Ic8EgPM08d_C5nEN60mxoRydlaQWySAemf_8Q_qVw,4175
156
- tpu_inference/runner/compilation_manager.py,sha256=pJFFLkFVmhXukBIxGRUo-hrOqx8jl8JIUuS36fZ2yvg,36177
159
+ tpu_inference/runner/compilation_manager.py,sha256=yIsonouB5G0-fyVtAKuyyRXaMGNFwnX8D7q6ppQYgUI,36318
157
160
  tpu_inference/runner/input_batch.py,sha256=bx221NX2IOWzrtopss-B-2ZKW4y-U6nQpG09PjpUziw,18273
158
- tpu_inference/runner/kv_cache.py,sha256=i54EbGQB-9bbOgk6KibTpJpTE2pfFuTfis7J1P_UB0M,4574
161
+ tpu_inference/runner/kv_cache.py,sha256=F4dzW2d53xuxkFUn0oKzwE6VklGUeVm-QM19NVfIQDU,4577
159
162
  tpu_inference/runner/kv_cache_manager.py,sha256=CJxXtdWuewJqcTBMoR70_Uvwxjtc3cK2jxe1KpI9kQc,22152
160
163
  tpu_inference/runner/lora_utils.py,sha256=B4xMCgXGJ4VNdePvn89HH3tIZ-gYsQ7Vq_YCiYIATEY,3843
161
164
  tpu_inference/runner/multimodal_manager.py,sha256=azEPdHOwz8CN11MQmorGdtrCLbFaTCxdWyuEsZTzjYM,9778
162
165
  tpu_inference/runner/persistent_batch_manager.py,sha256=KERSfKy6XjMejnbtPGI3hzoYAHJLeCxmpZVYPqBCago,11156
163
166
  tpu_inference/runner/speculative_decoding_manager.py,sha256=I3FDWKh2dn6nV8LgTGfCTwMKYnxQsTPpBIrmaJngXHs,10215
164
167
  tpu_inference/runner/structured_decoding_manager.py,sha256=Y0ERPhj4olFh6Y2TxP0R1_4UIJwy7nemYA-h63YIR2U,3622
165
- tpu_inference/runner/tpu_runner.py,sha256=5vPFey3KFnh5lczyj4cIT3mVhR8RuX8kbcuHVOg8DAg,72318
168
+ tpu_inference/runner/tpu_runner.py,sha256=3SZYn0CBA4LOaTO3GdQOxKx3HKmVcNmUEeSyzSAGyFY,73320
166
169
  tpu_inference/runner/utils.py,sha256=ZnWUoNo-7INeB0mdXti1jwUOdbmxyExznOs-crRTQLk,17126
167
170
  tpu_inference/spec_decode/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
168
171
  tpu_inference/spec_decode/jax/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
169
172
  tpu_inference/spec_decode/jax/eagle3.py,sha256=A1dt-dmBttpy-5DGcL4noEDCB0OGP8Xo6MXqgJvWIo8,16593
170
173
  tpu_inference/worker/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
171
- tpu_inference/worker/tpu_worker.py,sha256=KY7fH--NP7jiTduP5m0gDnmB2LbhIel0Ts37XmjYpPM,14207
172
- tpu_inference-0.11.1.dev202511130813.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
173
- tpu_inference-0.11.1.dev202511130813.dist-info/METADATA,sha256=LARdH4AAJfZrrU2Pj4EIN8Zl0QLjzEpzkRCqBbeUdT8,5465
174
- tpu_inference-0.11.1.dev202511130813.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
175
- tpu_inference-0.11.1.dev202511130813.dist-info/top_level.txt,sha256=gb1hRIQ3DOawUfVzvPL2E__2KPIl9I0vb5r0xcRBGYQ,20
176
- tpu_inference-0.11.1.dev202511130813.dist-info/RECORD,,
174
+ tpu_inference/worker/tpu_worker.py,sha256=0ZguK2BtIQjQSvyUTcUH9ENBrxt09w3CbgPoDY13Eok,14210
175
+ tpu_inference-0.11.1.dev202511180814.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
176
+ tpu_inference-0.11.1.dev202511180814.dist-info/METADATA,sha256=6dHy_ByQ0ihDNFuqyb-ZXTFczvQ8Ia54zBNTKaUPhSk,5465
177
+ tpu_inference-0.11.1.dev202511180814.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
178
+ tpu_inference-0.11.1.dev202511180814.dist-info/top_level.txt,sha256=gb1hRIQ3DOawUfVzvPL2E__2KPIl9I0vb5r0xcRBGYQ,20
179
+ tpu_inference-0.11.1.dev202511180814.dist-info/RECORD,,
/tpu_inference/layers/{jax → common}/binary_search.py RENAMED
File without changes
/tpu_inference/layers/{jax → common}/sharding.py RENAMED
File without changes