Commit Graph
23 Commits
Author SHA1 Message Date
qizixi c1e4a4052d [V1][Spec Decode] Support multi-layer eagle draft model (#18030)
Signed-off-by: qizixi <qizixi@meta.com>
2025-05-24 09:45:34 +00:00
qizixi d55e446d13 [V1][Spec Decode] Small refactors to improve eagle bookkeeping performance (#18424)
Signed-off-by: qizixi <qizixi@meta.com>
2025-05-24 06:51:22 +00:00
Jiayi YaoandRui Qiao 2628a69e35 [V1] Support Deepseek MTP (#18435)
Signed-off-by: Rui Qiao <ruisearch42@gmail.com>
Signed-off-by: YaoJiayi <120040070@link.cuhk.edu.cn>
Co-authored-by: Rui Qiao <ruisearch42@gmail.com>
2025-05-23 10:26:28 -07:00
Cyrus Leung fbb13a2c15 Revert "[V1] [Bugfix] eagle bugfix and enable correct lm_head for multimodal (#18034)" (#18600)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
2025-05-23 02:18:22 -07:00
RonaldBXu 4c611348a7 [V1] [Bugfix] eagle bugfix and enable correct lm_head for multimodal (#18034)
Signed-off-by: Ronald Xu <ronaldxu@amazon.com>
2025-05-23 00:37:18 -07:00
Mark McLoughlin c6b636f9fb [V1][Spec Decoding] Use model_loader.get_model() to load models (#18273)
Signed-off-by: Mark McLoughlin <markmc@redhat.com>
2025-05-23 02:05:44 +00:00
Ekagra Ranjan a04720bc36 [V1][Spec Decode][Bugfix] Load quantize weights for EAGLE (#18290) 2025-05-22 15:17:33 -07:00
418d2f8bfb [V1][Spec Decode] Share input embedding of target model with EAGLE draft model to free ~1GB for llama 3 model (#17326)
Co-authored-by: root <root@ekagra-8xh100.us-east5-a.c.serving-efficiency-poc.internal>
Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
2025-05-14 12:31:46 -07:00
wwl2755 dc9905368d [V1][Spec Decode] Eagle unit tests (#17350)
Signed-off-by: wwl2755 <wangwenlong2755@gmail.com>
2025-05-12 23:01:17 +00:00
Jee Jee Li 822de7fb94 [Misc] Split model loader (#17712)
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
2025-05-07 12:42:26 +08:00
Woosuk Kwon 8d84d836d1 [BugFix][Spec Decode] Fix hidden size mismatch between target and eagle head (#17740)
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
2025-05-06 19:51:26 -07:00
Chen Zhang cba31c47c4 [v1] AttentionMetadata for each layer (#17394)
Signed-off-by: Chen Zhang <zhangch99@outlook.com>
2025-05-06 07:58:37 -07:00
Mengqing Cao f9bc5a0693 [Bugfix] Fix triton import with local TritonPlaceholder (#17446)
Signed-off-by: Mengqing Cao <cmq0113@163.com>
2025-05-06 17:53:09 +08:00
qizixi 39c0813a7f [V1][Spec Decode] Apply torch.compile & cudagraph to EAGLE3 (#17504)
Signed-off-by: qizixi <qizixi@meta.com>
2025-05-01 16:19:30 -07:00
Bryan Lu 70788bdbdc [V1][Spec Decode] Apply torch.compile & cudagraph to EAGLE (#17211)
Signed-off-by: Bryan Lu <yuzhelu@amazon.com>
2025-04-29 21:10:00 +00:00
Ekagra Ranjan e136000595 [V1][Spec Decode] Make Eagle model arch config driven (#17323) 2025-04-29 10:22:02 +08:00
Benjamin ChislettandBryan Lu a0e619e62a [V1][Spec Decode] EAGLE-3 Support (#16937)
Signed-off-by: Bryan Lu <yuzhelu@amazon.com>
Signed-off-by: Benjamin Chislett <benjamin.chislett@centml.ai>
Co-authored-by: Bryan Lu <yuzhelu@amazon.com>
2025-04-25 15:43:07 -07:00
Woosuk Kwon 41fb013d29 [V1][Spec Decode] Always use argmax for sampling draft tokens (#16899)
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
2025-04-23 14:57:43 -07:00
Woosuk Kwon 1311913f55 [BugFix][Spec Decode] No in-place update to draft probs (#16952)
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
2025-04-21 19:54:19 -07:00
Woosuk Kwon 3a0fba5cf4 [V1][Spec Decode] Handle draft tokens beyond max_model_len (#16087)
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
2025-04-21 12:38:50 -07:00
Lily Liu e8224f3dca [V1][Spec Decode] Eagle Model loading (#16035)
Signed-off-by: LiuXiaoxuanPKU <lilyliupku@gmail.com>
2025-04-10 11:21:48 -07:00
Ekagra Ranjan 24b7fb455a [Spec Decode] Fix input triton kernel for eagle (#15909) 2025-04-01 18:15:14 -07:00
Woosuk Kwon e75a6301bd [V1][Spec Decode] Implement Eagle Proposer [1/N] (#15729)
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
2025-04-01 12:33:16 -07:00