optimize ulysses-4090: balanced a2a and merged qkv (#590)

Co-authored-by: root <root@pt-de4c35727a1b4d1b9f27f422f06026ec-worker-0.pt-de4c35727a1b4d1b9f27f422f06026ec.ns-devsft-3460edd0.svc.cluster.local>
Co-authored-by: root <root@pt-9b2035a55fe647eeb007584b238e5077-worker-0.pt-9b2035a55fe647eeb007584b238e5077.ns-devsft-3460edd0.svc.cluster.local>

configs/seko_talk/seko_talk_25_int8_dist_fp8_comm.json

@@ -4,9 +4,9 @@
     "video_duration": 5,
     "audio_sr": 16000,
     "target_video_length": 81,
-    "self_attn_1_type": "sage_attn3",
-    "cross_attn_1_type": "sage_attn3",
-    "cross_attn_2_type": "sage_attn3",
+    "self_attn_1_type": "sage_attn2",
+    "cross_attn_1_type": "sage_attn2",
+    "cross_attn_2_type": "sage_attn2",
     "sample_guide_scale": 1,
     "sample_shift": 5,
     "enable_cfg": false,
@@ -14,16 +14,16 @@
     "cpu_offload": false,
     "offload_granularity": "block",
     "offload_ratio": 1,
-    "t5_cpu_offload": false,
+    "t5_cpu_offload": true,
     "t5_quantized": true,
     "t5_quant_scheme": "int8-q8f",
-    "clip_cpu_offload": false,
+    "clip_cpu_offload": true,
     "clip_quantized": false,
-    "audio_encoder_cpu_offload": false,
-    "audio_adapter_cpu_offload": false,
+    "audio_encoder_cpu_offload": true,
+    "audio_adapter_cpu_offload": true,
     "adapter_quantized": true,
     "adapter_quant_scheme": "int8-q8f",
-    "vae_cpu_offload": false,
+    "vae_cpu_offload": true,
     "use_tiling_vae": false,
     "dit_quantized": true,
     "dit_quant_scheme": "int8-q8f",
@@ -35,6 +35,6 @@
     "parallel": {
         "seq_p_size": 8,
         "seq_p_fp8_comm": true,
-        "seq_p_attn_type": "ulysses"
+        "seq_p_attn_type": "ulysses-4090"
     }
 }