Initial commit

configs/seko_talk/5090/seko_talk_5090_bf16.json

+{
+    "infer_steps": 4,
+    "target_fps": 16,
+    "video_duration": 360,
+    "audio_sr": 16000,
+    "target_video_length": 81,
+    "resize_mode": "adaptive",
+    "self_attn_1_type": "sage_attn3",
+    "cross_attn_1_type": "sage_attn3",
+    "cross_attn_2_type": "sage_attn3",
+    "sample_guide_scale": 1,
+    "sample_shift": 5,
+    "enable_cfg": false,
+    "use_31_block": true,
+    "cpu_offload": true,
+    "offload_granularity": "block",
+    "offload_ratio": 1,
+    "t5_cpu_offload": true,
+    "clip_cpu_offload": false,
+    "audio_encoder_cpu_offload": false,
+    "audio_adapter_cpu_offload": false,
+    "vae_cpu_offload": false
+}

configs/seko_talk/5090/seko_talk_5090_int8.json

+{
+    "infer_steps": 4,
+    "target_fps": 16,
+    "video_duration": 360,
+    "audio_sr": 16000,
+    "target_video_length": 81,
+    "resize_mode": "adaptive",
+    "self_attn_1_type": "sage_attn3",
+    "cross_attn_1_type": "sage_attn3",
+    "cross_attn_2_type": "sage_attn3",
+    "sample_guide_scale": 1,
+    "sample_shift": 5,
+    "enable_cfg": false,
+    "use_31_block": true,
+    "cpu_offload": true,
+    "offload_granularity": "block",
+    "offload_ratio": 1,
+    "t5_cpu_offload": false,
+    "clip_cpu_offload": false,
+    "audio_encoder_cpu_offload": false,
+    "audio_adapter_cpu_offload": false,
+    "vae_cpu_offload": false,
+    "dit_quantized": true,
+    "dit_quant_scheme": "int8-q8f",
+    "adapter_quantized": true,
+    "adapter_quant_scheme": "int8-q8f",
+    "t5_quantized": true,
+    "t5_quant_scheme": "int8-q8f"
+}

configs/seko_talk/5090/seko_talk_5090_int8_8gpu.json

+
+{
+    "infer_steps": 4,
+    "target_fps": 16,
+    "video_duration": 360,
+    "audio_sr": 16000,
+    "target_video_length": 81,
+    "resize_mode": "adaptive",
+    "self_attn_1_type": "sage_attn3",
+    "cross_attn_1_type": "sage_attn3",
+    "cross_attn_2_type": "sage_attn3",
+    "sample_guide_scale": 1,
+    "sample_shift": 5,
+    "enable_cfg": false,
+    "use_31_block": true,
+    "cpu_offload": true,
+    "offload_granularity": "block",
+    "offload_ratio": 1,
+    "t5_cpu_offload": false,
+    "clip_cpu_offload": false,
+    "audio_encoder_cpu_offload": false,
+    "audio_adapter_cpu_offload": false,
+    "vae_cpu_offload": false,
+    "dit_quantized": true,
+    "dit_quant_scheme": "int8-q8f",
+    "adapter_quantized": true,
+    "adapter_quant_scheme": "int8-q8f",
+    "t5_quantized": true,
+    "t5_quant_scheme": "int8-q8f",
+    "parallel": {
+        "seq_p_size": 8,
+        "seq_p_attn_type": "ulysses-4090"
+    }
+}

configs/seko_talk/A800/seko_talk_A800_int8.json

+{
+    "infer_steps": 4,
+    "target_fps": 16,
+    "video_duration": 360,
+    "audio_sr": 16000,
+    "target_video_length": 81,
+    "resize_mode": "adaptive",
+    "self_attn_1_type": "flash_attn3",
+    "cross_attn_1_type": "flash_attn3",
+    "cross_attn_2_type": "flash_attn3",
+    "sample_guide_scale": 1.0,
+    "sample_shift": 5,
+    "enable_cfg": false,
+    "cpu_offload": false,
+    "use_31_block": true,
+    "dit_quantized": true,
+    "dit_quant_scheme": "int8-vllm",
+    "adapter_quantized": true,
+    "adapter_quant_scheme": "int8-vllm",
+    "t5_quantized": true,
+    "t5_quant_scheme": "int8-vllm"
+}

configs/seko_talk/A800/seko_talk_A800_int8_dist_2gpu.json

+{
+    "infer_steps": 4,
+    "target_fps": 16,
+    "video_duration": 360,
+    "audio_sr": 16000,
+    "target_video_length": 81,
+    "resize_mode": "adaptive",
+    "self_attn_1_type": "flash_attn3",
+    "cross_attn_1_type": "flash_attn3",
+    "cross_attn_2_type": "flash_attn3",
+    "sample_guide_scale": 1.0,
+    "sample_shift": 5,
+    "enable_cfg": false,
+    "cpu_offload": false,
+    "use_31_block": true,
+    "dit_quantized": true,
+    "dit_quant_scheme": "int8-vllm",
+    "adapter_quantized": true,
+    "adapter_quant_scheme": "int8-vllm",
+    "t5_quantized": true,
+    "t5_quant_scheme": "int8-vllm",
+    "parallel": {
+        "seq_p_size": 2,
+        "seq_p_attn_type": "ulysses"
+    }
+}

configs/seko_talk/A800/seko_talk_A800_int8_dist_4gpu.json

+{
+    "infer_steps": 4,
+    "target_fps": 16,
+    "video_duration": 360,
+    "audio_sr": 16000,
+    "target_video_length": 81,
+    "resize_mode": "adaptive",
+    "self_attn_1_type": "flash_attn3",
+    "cross_attn_1_type": "flash_attn3",
+    "cross_attn_2_type": "flash_attn3",
+    "sample_guide_scale": 1.0,
+    "sample_shift": 5,
+    "enable_cfg": false,
+    "cpu_offload": false,
+    "use_31_block": true,
+    "dit_quantized": true,
+    "dit_quant_scheme": "int8-vllm",
+    "adapter_quantized": true,
+    "adapter_quant_scheme": "int8-vllm",
+    "t5_quantized": true,
+    "t5_quant_scheme": "int8-vllm",
+    "parallel": {
+        "seq_p_size": 4,
+        "seq_p_attn_type": "ulysses"
+    }
+}

configs/seko_talk/A800/seko_talk_A800_int8_dist_8gpu.json

+{
+    "infer_steps": 4,
+    "target_fps": 16,
+    "video_duration": 360,
+    "audio_sr": 16000,
+    "target_video_length": 81,
+    "resize_mode": "adaptive",
+    "self_attn_1_type": "flash_attn3",
+    "cross_attn_1_type": "flash_attn3",
+    "cross_attn_2_type": "flash_attn3",
+    "sample_guide_scale": 1.0,
+    "sample_shift": 5,
+    "enable_cfg": false,
+    "cpu_offload": false,
+    "use_31_block": true,
+    "dit_quantized": true,
+    "dit_quant_scheme": "int8-vllm",
+    "adapter_quantized": true,
+    "adapter_quant_scheme": "int8-vllm",
+    "t5_quantized": true,
+    "t5_quant_scheme": "int8-vllm",
+    "parallel": {
+        "seq_p_size": 8,
+        "seq_p_attn_type": "ulysses"
+    }
+}

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

+{
+    "infer_steps": 4,
+    "target_fps": 16,
+    "video_duration": 360,
+    "audio_sr": 16000,
+    "target_video_length": 81,
+    "resize_mode": "adaptive",
+    "self_attn_1_type": "sage_attn2",
+    "cross_attn_1_type": "sage_attn2",
+    "cross_attn_2_type": "sage_attn2",
+    "sample_guide_scale": 1.0,
+    "sample_shift": 5,
+    "enable_cfg": false,
+    "use_31_block": true,
+    "cpu_offload": true,
+    "offload_granularity": "block",
+    "offload_ratio": 0.8,
+    "t5_cpu_offload": false,
+    "clip_cpu_offload": false,
+    "vae_cpu_offload": false,
+    "audio_encoder_cpu_offload": false,
+    "audio_adapter_cpu_offload": false
+}

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

+{
+    "infer_steps": 4,
+    "target_fps": 16,
+    "video_duration": 360,
+    "audio_sr": 16000,
+    "target_video_length": 81,
+    "resize_mode": "adaptive",
+    "self_attn_1_type": "sage_attn2",
+    "cross_attn_1_type": "sage_attn2",
+    "cross_attn_2_type": "sage_attn2",
+    "sample_guide_scale": 1.0,
+    "sample_shift": 5,
+    "enable_cfg": false,
+    "use_31_block": true,
+    "t5_quantized": true,
+    "t5_quant_scheme": "fp8-q8f",
+    "dit_quantized": true,
+    "dit_quant_scheme": "fp8-q8f",
+    "adapter_quantized": true,
+    "adapter_quant_scheme": "fp8",
+    "cpu_offload": false,
+    "t5_cpu_offload": true,
+    "clip_cpu_offload": true,
+    "vae_cpu_offload": true,
+    "audio_encoder_cpu_offload": true,
+    "audio_adapter_cpu_offload": true
+}

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

+{
+    "infer_steps": 4,
+    "target_fps": 16,
+    "video_duration": 360,
+    "audio_sr": 16000,
+    "target_video_length": 81,
+    "resize_mode": "adaptive",
+    "self_attn_1_type": "sage_attn2",
+    "cross_attn_1_type": "sage_attn2",
+    "cross_attn_2_type": "sage_attn2",
+    "sample_guide_scale": 1.0,
+    "sample_shift": 5,
+    "enable_cfg": false,
+    "use_31_block": true,
+    "cpu_offload": false,
+    "t5_cpu_offload": true,
+    "clip_cpu_offload": true,
+    "vae_cpu_offload": true,
+    "audio_encoder_cpu_offload": true,
+    "audio_adapter_cpu_offload": true,
+    "parallel": {
+        "seq_p_size": 2,
+        "seq_p_attn_type": "ulysses"
+    }
+}

configs/seko_talk/L40s/2gpu/seko_talk_fp8.json

+{
+    "infer_steps": 4,
+    "target_fps": 16,
+    "video_duration": 360,
+    "audio_sr": 16000,
+    "target_video_length": 81,
+    "resize_mode": "adaptive",
+    "self_attn_1_type": "sage_attn2",
+    "cross_attn_1_type": "sage_attn2",
+    "cross_attn_2_type": "sage_attn2",
+    "sample_guide_scale": 1.0,
+    "sample_shift": 5,
+    "enable_cfg": false,
+    "use_31_block": true,
+    "t5_quantized": true,
+    "t5_quant_scheme": "fp8-q8f",
+    "dit_quantized": true,
+    "dit_quant_scheme": "fp8-q8f",
+    "adapter_quantized": true,
+    "adapter_quant_scheme": "fp8",
+    "cpu_offload": false,
+    "t5_cpu_offload": true,
+    "clip_cpu_offload": true,
+    "vae_cpu_offload": true,
+    "audio_encoder_cpu_offload": true,
+    "audio_adapter_cpu_offload": true,
+    "parallel": {
+        "seq_p_size": 2,
+        "seq_p_attn_type": "ulysses"
+    }
+}

configs/seko_talk/L40s/4gpu/seko_talk_bf16.json

+{
+    "infer_steps": 4,
+    "target_fps": 16,
+    "video_duration": 360,
+    "audio_sr": 16000,
+    "target_video_length": 81,
+    "resize_mode": "adaptive",
+    "self_attn_1_type": "sage_attn2",
+    "cross_attn_1_type": "sage_attn2",
+    "cross_attn_2_type": "sage_attn2",
+    "sample_guide_scale": 1.0,
+    "sample_shift": 5,
+    "enable_cfg": false,
+    "use_31_block": true,
+    "cpu_offload": false,
+    "t5_cpu_offload": true,
+    "clip_cpu_offload": true,
+    "vae_cpu_offload": true,
+    "audio_encoder_cpu_offload": true,
+    "audio_adapter_cpu_offload": true,
+    "parallel": {
+        "seq_p_size": 4,
+        "seq_p_attn_type": "ulysses"
+    }
+}

configs/seko_talk/L40s/4gpu/seko_talk_fp8.json

+{
+    "infer_steps": 4,
+    "target_fps": 16,
+    "video_duration": 360,
+    "audio_sr": 16000,
+    "target_video_length": 81,
+    "resize_mode": "adaptive",
+    "self_attn_1_type": "sage_attn2",
+    "cross_attn_1_type": "sage_attn2",
+    "cross_attn_2_type": "sage_attn2",
+    "sample_guide_scale": 1.0,
+    "sample_shift": 5,
+    "enable_cfg": false,
+    "use_31_block": true,
+    "t5_quantized": true,
+    "t5_quant_scheme": "fp8-q8f",
+    "dit_quantized": true,
+    "dit_quant_scheme": "fp8-q8f",
+    "adapter_quantized": true,
+    "adapter_quant_scheme": "fp8",
+    "cpu_offload": false,
+    "t5_cpu_offload": true,
+    "clip_cpu_offload": true,
+    "vae_cpu_offload": true,
+    "audio_encoder_cpu_offload": true,
+    "audio_adapter_cpu_offload": true,
+    "parallel": {
+        "seq_p_size": 4,
+        "seq_p_attn_type": "ulysses"
+    }
+}

configs/seko_talk/L40s/8gpu/seko_talk_bf16.json

+{
+    "infer_steps": 4,
+    "target_fps": 16,
+    "video_duration": 360,
+    "audio_sr": 16000,
+    "target_video_length": 81,
+    "resize_mode": "adaptive",
+    "self_attn_1_type": "sage_attn2",
+    "cross_attn_1_type": "sage_attn2",
+    "cross_attn_2_type": "sage_attn2",
+    "sample_guide_scale": 1.0,
+    "sample_shift": 5,
+    "enable_cfg": false,
+    "use_31_block": true,
+    "cpu_offload": false,
+    "t5_cpu_offload": true,
+    "clip_cpu_offload": true,
+    "vae_cpu_offload": true,
+    "audio_encoder_cpu_offload": true,
+    "audio_adapter_cpu_offload": true,
+    "parallel": {
+        "seq_p_size": 8,
+        "seq_p_attn_type": "ulysses"
+    }
+}

configs/seko_talk/L40s/8gpu/seko_talk_fp8.json

+{
+    "infer_steps": 4,
+    "target_fps": 16,
+    "video_duration": 360,
+    "audio_sr": 16000,
+    "target_video_length": 81,
+    "resize_mode": "adaptive",
+    "self_attn_1_type": "sage_attn2",
+    "cross_attn_1_type": "sage_attn2",
+    "cross_attn_2_type": "sage_attn2",
+    "sample_guide_scale": 1.0,
+    "sample_shift": 5,
+    "enable_cfg": false,
+    "use_31_block": true,
+    "t5_quantized": true,
+    "t5_quant_scheme": "fp8-q8f",
+    "dit_quantized": true,
+    "dit_quant_scheme": "fp8-q8f",
+    "adapter_quantized": true,
+    "adapter_quant_scheme": "fp8",
+    "cpu_offload": false,
+    "t5_cpu_offload": true,
+    "clip_cpu_offload": true,
+    "vae_cpu_offload": true,
+    "audio_encoder_cpu_offload": true,
+    "audio_adapter_cpu_offload": true,
+    "parallel": {
+        "seq_p_size": 8,
+        "seq_p_attn_type": "ulysses"
+    }
+}

configs/seko_talk/mlu/seko_talk_bf16.json

+{
+    "infer_steps": 4,
+    "target_fps": 16,
+    "video_duration": 360,
+    "audio_sr": 16000,
+    "target_video_length": 81,
+    "resize_mode": "adaptive",
+    "self_attn_1_type": "mlu_sage_attn",
+    "cross_attn_1_type": "mlu_sage_attn",
+    "cross_attn_2_type": "mlu_sage_attn",
+    "sample_guide_scale": 1.0,
+    "sample_shift": 5,
+    "enable_cfg": false,
+    "cpu_offload": false,
+    "use_31_block": true,
+    "rope_type": "torch",
+    "modulate_type": "torch"
+}

configs/seko_talk/mlu/seko_talk_int8.json

+{
+    "infer_steps": 4,
+    "target_fps": 16,
+    "video_duration": 360,
+    "audio_sr": 16000,
+    "target_video_length": 81,
+    "resize_mode": "adaptive",
+    "self_attn_1_type": "mlu_sage_attn",
+    "cross_attn_1_type": "mlu_sage_attn",
+    "cross_attn_2_type": "mlu_sage_attn",
+    "seed": 42,
+    "sample_guide_scale": 1.0,
+    "sample_shift": 5,
+    "enable_cfg": false,
+    "cpu_offload": false,
+    "use_31_block": true,
+    "clip_quantized": false,
+    "clip_quant_scheme": "int8-tmo",
+    "dit_quantized": true,
+    "dit_quant_scheme": "int8-tmo",
+    "adapter_quantized": true,
+    "adapter_quant_scheme": "int8-tmo",
+    "t5_quantized": true,
+    "t5_quant_scheme": "int8-tmo",
+    "modulate_type": "torch",
+    "rope_type": "torch",
+    "layer_norm_type": "torch",
+    "rms_norm_type": "torch"
+}

configs/seko_talk/mlu/seko_talk_int8_dist.json

+{
+    "infer_steps": 4,
+    "target_fps": 16,
+    "video_duration": 360,
+    "audio_sr": 16000,
+    "target_video_length": 81,
+    "resize_mode": "adaptive",
+    "self_attn_1_type": "mlu_sage_attn",
+    "cross_attn_1_type": "mlu_sage_attn",
+    "cross_attn_2_type": "mlu_sage_attn",
+    "seed": 42,
+    "sample_guide_scale": 1.0,
+    "sample_shift": 5,
+    "enable_cfg": false,
+    "cpu_offload": false,
+    "use_31_block": true,
+    "clip_quantized": false,
+    "clip_quant_scheme": "int8-tmo",
+    "dit_quantized": true,
+    "dit_quant_scheme": "int8-tmo",
+    "adapter_quantized": true,
+    "adapter_quant_scheme": "int8-tmo",
+    "t5_quantized": true,
+    "t5_quant_scheme": "int8-tmo",
+    "modulate_type": "torch",
+    "rope_type": "torch",
+    "layer_norm_type": "torch",
+    "rms_norm_type": "torch",
+    "parallel": {
+        "seq_p_size": 8,
+        "seq_p_attn_type": "ulysses"
+    }
+}

configs/seko_talk/multi_person/01_base.json

+{
+    "infer_steps": 4,
+    "target_fps": 16,
+    "video_duration": 360,
+    "audio_sr": 16000,
+    "target_video_length": 81,
+    "resize_mode": "adaptive",
+    "self_attn_1_type": "flash_attn3",
+    "cross_attn_1_type": "flash_attn3",
+    "cross_attn_2_type": "flash_attn3",
+    "sample_guide_scale": 1.0,
+    "sample_shift": 5,
+    "enable_cfg": false,
+    "cpu_offload": false,
+    "use_31_block": true
+}

configs/seko_talk/multi_person/02_base_fp8.json

+{
+    "infer_steps": 4,
+    "target_fps": 16,
+    "video_duration": 360,
+    "audio_sr": 16000,
+    "target_video_length": 81,
+    "resize_mode": "adaptive",
+    "self_attn_1_type": "sage_attn2",
+    "cross_attn_1_type": "sage_attn2",
+    "cross_attn_2_type": "sage_attn2",
+    "sample_guide_scale": 1.0,
+    "sample_shift": 5,
+    "enable_cfg": false,
+    "cpu_offload": false,
+    "use_31_block": true,
+    "dit_quantized": true,
+    "dit_quant_scheme": "fp8-sgl",
+    "adapter_quantized": true,
+    "adapter_quant_scheme": "fp8",
+    "t5_quantized": true,
+    "t5_quant_scheme": "fp8"
+}