File tree Expand file tree Collapse file tree 3 files changed +0
-85
lines changed Expand file tree Collapse file tree 3 files changed +0
-85
lines changed Original file line number Diff line number Diff line change 105
105
# Future Plan:
106
106
# Revert it when the related pr is merged in vllm and vllm-ascend.
107
107
#
108
- # ** File: worker/patch_common/patch_eagle.py **
109
- # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
110
- # 1. `vllm.v1.spec_decode.eagle.prepare_inputs`
111
- # Why:
112
- # We need to use the patched `prepare_input_kernel` in `eagle.prepare_inputs`.
113
- # The mainly reason to overwrite `prepare_input_kernel` is this is a triton
114
- # kernel, ascend is now not support triton kernel.
115
- # How:
116
- # Re-implementation the `prepare_input_kernel` triton kernel by pytorch
117
- # Related PR (if no, explain why):
118
- # - Ascend doesn't support triton
119
- # Future Plan:
120
- # Revert it when the ascend support triton kernel.
121
- #
122
108
# ** File: worker/patch_common/patch_sampler.py **
123
109
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
124
110
# 1. `vllm.v1.sample.sampler.Sampler.apply_top_k_top_p`
Original file line number Diff line number Diff line change 19
19
# patch files.
20
20
import vllm_ascend .patch .worker .patch_common .patch_utils # noqa isort:skip
21
21
import vllm_ascend .patch .worker .patch_common .patch_distributed # noqa
22
- import vllm_ascend .patch .worker .patch_common .patch_eagle # noqa
23
22
import vllm_ascend .patch .worker .patch_common .patch_minicpm # noqa
24
23
import vllm_ascend .patch .worker .patch_common .patch_multi_step_worker # noqa
25
24
import vllm_ascend .patch .worker .patch_common .patch_sampler # noqa
Load Diff This file was deleted.
You can’t perform that action at this time.
0 commit comments