vllm.v1.kv_offload.tiering.fs.thread_pool ¶
Thread pool
Two queues (load, store) and two sets of threads: - Load-priority threads: drain the load queue first, then the store queue. - Store-priority threads: drain the store queue first, then the load queue. Load jobs are enqueued to the load queue; store jobs to the store queue.
DualQueueThreadPool ¶
Thread pool with two task queues (load and store) and two thread groups.
Load-priority threads drain the load queue first, then fall back to the store queue. Store-priority threads do the reverse. Both queues share a single condition variable.
Source code in vllm/v1/kv_offload/tiering/fs/thread_pool.py
50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 | |
enqueue_load ¶
Enqueue load tasks for a job (high-priority for load-priority threads).
Source code in vllm/v1/kv_offload/tiering/fs/thread_pool.py
enqueue_store ¶
Enqueue store tasks for a job (high-priority for store-priority threads).
Source code in vllm/v1/kv_offload/tiering/fs/thread_pool.py
JobState ¶
Thread-safe completion tracker for a set of per-block I/O tasks.
Each task calls task_done(success) when it finishes.
Source code in vllm/v1/kv_offload/tiering/fs/thread_pool.py
task_done ¶
Returns if job completed and success flag