w8a8-int8 version

jakep-allenai · jakep-allenai · commit 022f437b41f9 · 2025-07-15T21:02:35.000Z
diff --git a/olmocr/train/quantization_configs/qwen2_5vl_w8a8_int8.yaml b/olmocr/train/quantization_configs/qwen2_5vl_w8a8_int8.yaml
@@ -0,0 +1,8 @@
+quantization_stage:
+  quantization_modifiers:
+    GPTQModifier:
+      targets: [Linear]
+      ignore: ['re:.*lm_head', 're:model.visual.*']
+      weights:
+        num_bits: 8
+        type: int
