Merge pull request #394 from ModelTC/vlm

fastvid

Author: gushiqiao

configs/sparsification/methods/FastVID/fastvid.yml

@@ -0,0 +1,28 @@
+base:
+    seed: &seed 42
+model:
+    type: Llava OneVision
+    path: model path
+    torch_dtype: auto
+eval:
+    eval_pos: [pretrain, transformed]
+    type: vqa
+    name: [mme]
+    download: False
+    path: MME dataset path
+    bs: 1
+    inference_per_block: False
+sparse:
+    method: TokenReduction
+    special:
+        method: FastVID
+        retention_ratio: 0.10
+        DySeg_c: 8
+        DySeg_tau: 0.9
+        STPrune_d: 0.4
+        DTM_p: 4
+        DTM_alpha: 0.6
+save:
+    save_trans: False
+    save_fake: False
+    save_path: /path/to/save/

llmc/compression/token_reduction/__init__.py

@@ -2,6 +2,7 @@
 from .dycoke import DyCoke
 from .fastervlm import FasterVLM
 from .fastv import FastV
+from .fastvid import FastVID
 from .prunevid import PruneVid
 from .pyramiddrop import PyramidDrop
 from .sparsevlm import SparseVLM