Dev/seko talk v2.7 (#867)

Co-authored-by: gaclove <peng.gaoc@gmail.com>

Author: wangshankun

configs/seko_talk/5090/seko_talk_5090_bf16.json

@@ -11,7 +11,7 @@
     "sample_guide_scale": 1,
     "sample_shift": 5,
     "enable_cfg": false,
-    "use_31_block": false,
+    "use_31_block": true,
     "cpu_offload": true,
     "offload_granularity": "block",
     "offload_ratio": 1,

configs/seko_talk/5090/seko_talk_5090_int8.json

@@ -11,7 +11,7 @@
     "sample_guide_scale": 1,
     "sample_shift": 5,
     "enable_cfg": false,
-    "use_31_block": false,
+    "use_31_block": true,
     "cpu_offload": true,
     "offload_granularity": "block",
     "offload_ratio": 1,

configs/seko_talk/5090/seko_talk_5090_int8_8gpu.json

@@ -12,7 +12,7 @@
     "sample_guide_scale": 1,
     "sample_shift": 5,
     "enable_cfg": false,
-    "use_31_block": false,
+    "use_31_block": true,
     "cpu_offload": true,
     "offload_granularity": "block",
     "offload_ratio": 1,

configs/seko_talk/A800/seko_talk_A800_int8.json

@@ -12,7 +12,7 @@
     "sample_shift": 5,
     "enable_cfg": false,
     "cpu_offload": false,
-    "use_31_block": false,
+    "use_31_block": true,
     "dit_quantized": true,
     "dit_quant_scheme": "int8-vllm",
     "adapter_quantized": true,

configs/seko_talk/A800/seko_talk_A800_int8_dist_2gpu.json

@@ -12,7 +12,7 @@
     "sample_shift": 5,
     "enable_cfg": false,
     "cpu_offload": false,
-    "use_31_block": false,
+    "use_31_block": true,
     "dit_quantized": true,
     "dit_quant_scheme": "int8-vllm",
     "adapter_quantized": true,

configs/seko_talk/A800/seko_talk_A800_int8_dist_4gpu.json

@@ -12,7 +12,7 @@
     "sample_shift": 5,
     "enable_cfg": false,
     "cpu_offload": false,
-    "use_31_block": false,
+    "use_31_block": true,
     "dit_quantized": true,
     "dit_quant_scheme": "int8-vllm",
     "adapter_quantized": true,

configs/seko_talk/A800/seko_talk_A800_int8_dist_8gpu.json

@@ -12,7 +12,7 @@
     "sample_shift": 5,
     "enable_cfg": false,
     "cpu_offload": false,
-    "use_31_block": false,
+    "use_31_block": true,
     "dit_quantized": true,
     "dit_quant_scheme": "int8-vllm",
     "adapter_quantized": true,

configs/seko_talk/L40s/1gpu/seko_talk_bf16.json

@@ -11,7 +11,7 @@
     "sample_guide_scale": 1.0,
     "sample_shift": 5,
     "enable_cfg": false,
-    "use_31_block": false,
+    "use_31_block": true,
     "cpu_offload": true,
     "offload_granularity": "block",
     "offload_ratio": 0.8,

configs/seko_talk/L40s/1gpu/seko_talk_fp8.json

@@ -11,7 +11,7 @@
     "sample_guide_scale": 1.0,
     "sample_shift": 5,
     "enable_cfg": false,
-    "use_31_block": false,
+    "use_31_block": true,
     "t5_quantized": true,
     "t5_quant_scheme": "fp8-q8f",
     "dit_quantized": true,

configs/seko_talk/L40s/2gpu/seko_talk_bf16.json

@@ -11,7 +11,7 @@
     "sample_guide_scale": 1.0,
     "sample_shift": 5,
     "enable_cfg": false,
-    "use_31_block": false,
+    "use_31_block": true,
     "cpu_offload": false,
     "t5_cpu_offload": true,
     "clip_cpu_offload": true,