chore: update wan i2v/t2v configs and eval defaults

Made-with: Cursor

Author: Charles2530

configs/quantization/video_gen/wan_i2v/awq_w_a.yaml

@@ -2,7 +2,7 @@ base:
     seed: &seed 42
 model:
     type: WanI2V
-    path: /mnt/lm_data_afs/wangzining/charles/lab/llmc/models/Wan2.2-T2V-A14B/
+    path: /path/to/model
     torch_dtype: auto
 calib:
     name: i2v
@@ -31,12 +31,12 @@ quant:
     video_gen:
         method: Awq
         weight:
-            bit: 4
+            bit: 8
             symmetric: True
             granularity: per_channel
             group_size: -1
         act:
-            bit: 4
+            bit: 8
             symmetric: True
             granularity: per_token
         special:
@@ -46,4 +46,4 @@ quant:
             clip_sym: True
 save:
     save_lightx2v: True
-    save_path: ../lightx2v/wan_i2v_awq_w_a/x2v/
+    save_path: /path/to/x2v/

configs/quantization/video_gen/wan_i2v/smoothquant_w_a_fp8_example.yaml

@@ -0,0 +1,32 @@
+base:
+    seed: &seed 42
+model:
+    type: WanT2V
+    path: /path/to/wan_t2v
+    torch_dtype: auto
+eval:
+    eval_pos: [transformed, fake_quant]
+    type: video_gen
+    name: t2v
+    download: False
+    path: ../assets/wan_t2v/eval/
+    bs: 1
+    target_height: 480
+    target_width: 832
+    num_frames: 81
+    guidance_scale: 5.0
+    output_video_path: ./output_videos_rtn/
+quant:
+    video_gen:
+        method: RTN
+        weight:
+            bit: 6
+            symmetric: True
+            granularity: per_channel
+        act:
+            bit: 6
+            symmetric: True
+            granularity: per_token
+save:
+    save_lightx2v: True
+    save_path: /path/to/x2v/

configs/quantization/video_gen/wan_t2v/rtn_w_a.yaml

@@ -32,7 +32,7 @@ def main(config):
     logger.info(f'tokenizer: {model.get_tokenizer()}')
 
     eval_list = get_eval_list(model, config)
-    # eval_model(model, None, eval_list, eval_pos='pretrain')
+    eval_model(model, None, eval_list, eval_pos='pretrain')
 
     blockwise_opts = []
     modalities, modality_configs = get_modality(config)
@@ -70,7 +70,7 @@ def main(config):
             blockwise_opts.append(blockwise_opt)
             dist.barrier()
 
-    # eval_model(model, blockwise_opts, eval_list, eval_pos='transformed')
+    eval_model(model, blockwise_opts, eval_list, eval_pos='transformed')
     if int(os.environ['RANK']) == 0:
         if 'save' in config and config.save.get('save_trans', False):
             blockwise_opt.save_model(save_trans_path)
@@ -85,8 +85,8 @@ def main(config):
                 config.save.get('trtllm_cfg'),
             )
 
-        # eval_model(model, blockwise_opts, eval_list, eval_pos='fake_quant')
-        # eval_model(model, blockwise_opts, eval_list, eval_pos='fake_quant_wo_kv')
+        eval_model(model, blockwise_opts, eval_list, eval_pos='fake_quant')
+        eval_model(model, blockwise_opts, eval_list, eval_pos='fake_quant_wo_kv')
 
         if 'save' in config and config.save.get('save_fake', False):
             deploy_all_modality(blockwise_opts, 'fake_quant')