Add files using upload-large-folder tool

.download.log

@@ -0,0 +1,2 @@
+START 2026-08-01T00:40:47+08:00
+
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 
                                                                                                 

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/config.json

@@ -0,0 +1,356 @@
+{
+  "_gradient_checkpointing": true,
+  "ar_state_text_system_prompt": "You are a helpful assistant, capable of perceiving auditory and textual inputs, as well as generating text responses.",
+  "ar_target_debug": true,
+  "ar_target_do_sample": true,
+  "ar_target_max_new_tokens": 64,
+  "ar_target_temperature": 0.9,
+  "ar_target_top_k": 50,
+  "ar_target_top_p": 0.9,
+  "ar_target_use_v2": true,
+  "architectures": [
+    "Stage4DMDSelfRolloutMetaQuery"
+  ],
+  "attn_implementation": "auto",
+  "audio_context_noise": 0.5,
+  "audio_lora_dict": {
+    "adapter_name": "s2_ovi_audio_lora",
+    "enable": true,
+    "lora_dropout": 0.0,
+    "network_alpha": 128,
+    "rank": 128,
+    "target_modules": [
+      "q",
+      "k",
+      "v",
+      "o",
+      "ffn.0",
+      "ffn.2"
+    ]
+  },
+  "audio_loss_weight": 0.5,
+  "audio_negative_prompt": "robotic, muffled, echo, distorted",
+  "audio_vae_path": "./pretrained_models/MMAudio/ext_weights/",
+  "block_causal": {
+    "a2a_lookahead": -1,
+    "attention_backend": "sdpa",
+    "audio_kv_cache_size": 4,
+    "enable": true,
+    "frame_per_block": 3,
+    "teacher_forcing": true,
+    "v2a_lookahead": 1,
+    "video_kv_cache_size": 4
+  },
+  "bsmntw_weighting": true,
+  "connector_num_hidden_layers": 8,
+  "diffusion_forcing": {
+    "enable": false,
+    "prefix_only": false,
+    "prob": 0.9,
+    "start_step": 0,
+    "temp_chunk": 3
+  },
+  "diffusion_model_id": "Lightricks/LTX-Video",
+  "diffusion_model_path": "./pretrained_models/Ovi/model_960x960_10s.safetensors",
+  "distillation": {},
+  "dmd": {
+    "audio_loss_weight": 0.5,
+    "critic_audio_loss_weight": 0.5,
+    "critic_video_loss_weight": 1.0,
+    "denoising_step_lists": {
+      "1": [
+        1000
+      ],
+      "2": [
+        1000,
+        500
+      ],
+      "3": [
+        1000,
+        750,
+        500
+      ],
+      "4": [
+        1000,
+        750,
+        500,
+        250
+      ]
+    },
+    "denoising_step_lists_first_chunk": {
+      "1": [
+        1000,
+        750,
+        500,
+        250
+      ],
+      "2": [
+        1000,
+        750,
+        500,
+        250
+      ],
+      "3": [
+        1000,
+        750,
+        500,
+        250
+      ],
+      "4": [
+        1000,
+        750,
+        500,
+        250
+      ]
+    },
+    "detach_fusion_audio": false,
+    "detach_fusion_video": false,
+    "detach_mllm_mq": true,
+    "ema_start_step": 200,
+    "ema_weight": 0.99,
+    "fake_guidance_scale": 0.0,
+    "fake_score_learning_rate": 5e-06,
+    "fake_score_update_interval": 5,
+    "flowmatch_shift": 5.0,
+    "gan_audio_start_step": 0,
+    "gan_critic_checkpoint_blocks": 30,
+    "gan_feature_depths": [
+      0.1,
+      0.5,
+      0.9
+    ],
+    "gan_loss_weight_audio": 0.0,
+    "gan_loss_weight_video": 0.0,
+    "gan_t_max": 1000,
+    "gan_t_min": 0,
+    "gan_video_start_step": 0,
+    "generator_adam_beta1": 0.0,
+    "generator_audio_loss_weight": 0.5,
+    "generator_gt_audio_listening_weight_multiplier": 1.0,
+    "generator_gt_audio_loss_weight": 0.05,
+    "generator_gt_video_loss_weight": 0.0,
+    "generator_video_loss_weight": 1.0,
+    "gt_score_regularize_start_step": 93000,
+    "gt_score_regularize_weight": 0.5,
+    "lookahead_weight": 2.0,
+    "lookahead_weight_position": 1,
+    "min_score_timestep": 0,
+    "normalize_gradient": true,
+    "num_inference_steps": 2,
+    "num_train_timestep": 1000,
+    "plan_commit_consistency_enable": false,
+    "plan_commit_gt_weight": 1.0,
+    "plan_commit_loss_weight": 3.0,
+    "purify_draft_static_margin": 4.0,
+    "purify_draft_static_mask_enable": false,
+    "purify_draft_static_quantile": 0.2,
+    "purify_draft_weight": 0.0,
+    "purify_start_step": 0,
+    "purify_t_max": 100,
+    "purify_t_min": 0,
+    "purify_weight": 0.0,
+    "real_audio_scale": 4.5,
+    "real_guidance_scale": 3.5,
+    "real_score_gt_x0_ratio": 0.5,
+    "real_score_gt_x0_start_step": 36000,
+    "real_slg_layer": 11,
+    "score_flowmatch_shift": 5.0,
+    "trainer": "dmd_self_rollout",
+    "ts_schedule": false,
+    "ts_schedule_max": false,
+    "validation_score_timestep": 250,
+    "video_loss_weight": 1.0,
+    "warp_denoising_step": true
+  },
+  "fake_score_learning_rate": 5e-06,
+  "fake_score_lora_checkpointing": {
+    "audio_adapter_name": "s4_fake_ovi_audio_lora",
+    "block_causal": {
+      "enable": false
+    },
+    "merge_from_checkpoint": "./output/0217_s1_qwen2p5omni7b_ovi1p1_mq_nonverbal_context_and_T5_response_talking80_vision_audio_loss_0p5_ICL_voice_prefix_sink_loss",
+    "merge_load_non_lora": true,
+    "resume_from_checkpoint": null,
+    "resume_gan_head": false,
+    "resume_lora_only": false,
+    "video_adapter_name": "s4_fake_ovi_vid_lora"
+  },
+  "flowmatch_shift": 5.0,
+  "gen_fps": 25.0,
+  "generator_lora_checkpointing": {
+    "audio_adapter_name": "s2_ovi_audio_lora",
+    "merge_from_checkpoint": "./output/0217_s1_qwen2p5omni7b_ovi1p1_mq_nonverbal_context_and_T5_response_talking80_vision_audio_loss_0p5_ICL_voice_prefix_sink_loss",
+    "merge_load_non_lora": true,
+    "prefer_ema": true,
+    "resume_from_checkpoint": "./output/0222_stage3_causal_cd_frame3",
+    "resume_lora_only": true,
+    "video_adapter_name": "s2_ovi_vid_lora"
+  },
+  "has_context_prompt": {
+    "enable": true,
+    "gt_t5_transcript_ratio": 1.0,
+    "no_mq_context_uncond_ratio": 0.0,
+    "skip_mllm_encode_ratio": 0.3
+  },
+  "in_channels": 16,
+  "infer": false,
+  "latent_spatial_size": 28,
+  "latent_temporal_size": 25,
+  "listening_loss_weight": 1.0,
+  "listening_prompt": "not speaking, closed mouth, listening, silent",
+  "local_files_only": true,
+  "lora_checkpointing": {
+    "audio_adapter_name": "s2_ovi_audio_lora",
+    "merge_from_checkpoint": "./output/0217_s1_qwen2p5omni7b_ovi1p1_mq_nonverbal_context_and_T5_response_talking80_vision_audio_loss_0p5_ICL_voice_prefix_sink_loss",
+    "merge_load_non_lora": true,
+    "prefer_ema": true,
+    "resume_from_checkpoint": "./output/0222_stage3_causal_cd_frame3",
+    "resume_lora_only": true,
+    "video_adapter_name": "s2_ovi_vid_lora"
+  },
+  "lora_dict": {},
+  "max_input_text_tokens": 256,
+  "mllm_ce_only": false,
+  "mllm_id": "./pretrained_models/Qwen2.5-Omni-7B",
+  "mllm_local_path": null,
+  "mllm_lora_dict": {
+    "alpha": 64,
+    "dropout": 0.0,
+    "enable": false,
+    "rank": 32,
+    "target_modules": [
+      "q_proj",
+      "k_proj",
+      "v_proj",
+      "o_proj",
+      "gate_proj",
+      "up_proj",
+      "down_proj"
+    ]
+  },
+  "model_type": "stage4_dmd_self_rollout_metaquery",
+  "modules_to_freeze": [
+    "vae_model_video",
+    "vae_model_audio",
+    "text_model",
+    "fusion_model",
+    "mllm.mllm_backbone",
+    "mllm.connector"
+  ],
+  "modules_to_unfreeze": [],
+  "negative_text_prompt": "",
+  "noise_scheduler_id": "Lightricks/LTX-Video",
+  "num_metaqueries": 128,
+  "ovi_gradient_checkpointing_block_count": null,
+  "prev_temp_chunk": 1,
+  "real_score_lora_checkpointing": {
+    "audio_adapter_name": "s4_real_ovi_audio_lora",
+    "block_causal": {
+      "enable": false
+    },
+    "merge_from_checkpoint": "./output/0217_s1_qwen2p5omni7b_ovi1p1_mq_nonverbal_context_and_T5_response_talking80_vision_audio_loss_0p5_ICL_voice_prefix_sink_loss",
+    "merge_load_non_lora": true,
+    "resume_from_checkpoint": null,
+    "resume_lora_only": false,
+    "video_adapter_name": "s4_real_ovi_vid_lora"
+  },
+  "rollout": {
+    "agent_context_source": "gt",
+    "aud_kv_gt_ratio": 0.0,
+    "audio_exit_replay_grad": false,
+    "audio_kv_cache_size": 4,
+    "audio_kv_grad_blocks": 0,
+    "context_noise": 0,
+    "decode_context_min_video_latents": 12,
+    "enable_ode_step": false,
+    "exit_index_sync": "broadcast",
+    "extra_kv_clean_forward": false,
+    "kv_cache_size": 4,
+    "kv_shift_step": 0,
+    "last_step_only": false,
+    "listening_output_fade_samples": 1536,
+    "listening_silence_enable": false,
+    "num_windows_schedule": [
+      [
+        0,
+        [
+          [
+            1
+          ],
+          [
+            1
+          ]
+        ]
+      ],
+      [
+        500,
+        [
+          [
+            1,
+            2,
+            3
+          ],
+          [
+            0.4,
+            0.4,
+            0.2
+          ]
+        ]
+      ]
+    ],
+    "real_score_bridge_use_gt": false,
+    "same_step_across_blocks": true,
+    "score_bridge_vid_latents": 5,
+    "share_video_noise_weight": 0,
+    "use_kv_cache": true,
+    "user_context_source": "gt",
+    "validation_decode": true,
+    "validation_max_samples": 1,
+    "validation_num_windows": 3,
+    "validation_seed": 920,
+    "vid_kv_gt_ratio": 0.0,
+    "video_kv_cache_size": 4,
+    "video_latents_per_window": 24,
+    "window_duration_sec": 3.84
+  },
+  "scheduler_id": "Lightricks/LTX-Video",
+  "stage": 4,
+  "stage_name": "dmd_self_rollout",
+  "student_checkpoint": null,
+  "student_lora_checkpointing": {},
+  "system_prompt": "You are a speech-video clip narrator for a text-to-audio-video (T2AV) model.\nThe input includes audio-visual dialogue between the Interlocutor and the Agent. The dialogue context is provided ONLY to infer what occurs next and must NOT be described, referenced, or summarized explicitly.\nYour task is to write a narration describing the NEXT recorded speech and video clip that shows ONLY the AGENT. Follow these rules precisely:\nVISUAL RULES: - Provide a HIGH-DETAIL appearance sentence, covering: (a) Gender + precise clothing (material/color/style, e.g., \"velvet head covering with ruffled accent\"), (b) Distinctive features (eyeliner/beard/accessories), (c) Scene composition (framing/lighting/color tone/backdrop, e.g., \"stark black and white lighting against plain curtains\"). - Describe ONLY visible physical actions in strict chronological order: (a) Facial muscles (e.g., \"lips curve upward forming a subtle smile\"), (b) Eye movements (e.g., \"gaze drifts downward for two seconds\"), (c) Posture/gestures (e.g., \"thumb brushes beard thoughtfully\"). - Emotions MUST derive from observable cues: CORRECT: \"eyes crinkling at the corners in a warm smile\" (visible) WRONG: \"she feels happy\" (internal) or \"looks thoughtful\" (abstract).\nSPEECH RULES: - Determine speech occurrence based on PROVIDED TARGETS or NATURAL CONVERSATION FLOW: - IF TARGETS ARE PROVIDED: (a) If \"[Target State]: Listening\" is provided in the context, describe silent actions ONLY without any speech content and do not use <S><E> tags. (b) If \"[Target Speech]: [Text]\" is provided, use the provided [Text] exactly as full speech content. - IF TARGETS ARE NOT PROVIDED: (a) If dialogue context implies continued listening or non-verbal reaction, describe silent actions ONLY without any speech content and do not use <S><E> tags. (b) If dialogue context implies the agent WILL SPEAK NEXT (even after brief listening actions), generate speech. (c) When speech needs to be generated: (1) Infer the speech content EXCLUSIVELY from dialogue context. (2) The speech content must be a RESPONSE that move forward the conversation, but word overlap with the end of the agent\u2019s speech context is STRICTLY FORBIDDEN. - GENERAL SPEECH STYLE (Only for speech existing in the narration): (b) Wrap each segment in <S> and <E> (speech contents are broken down into multiple segments when there are too many words for one segment) in the MIDDLE of the narration. (c) For each speech segment, embed it NATURALLY within action descriptions BEFORE and AFTER speech: CORRECT: \"meeting the camera, <S>I have considered your proposal<E> he brushes his thumb against his beard.\" WRONG: separate speech/action blocks.\nAUDIO RULES: - Add \"Audio: [description]\" at the VERY END of the paragraph, separated by a single space. - If no speech occurs (means no <S><E> tags exist), use \"SILENT\" for the audio description. - Otherwise, describe ONLY the agent\u2019s audio characteristics: (a) Voice traits: Describe the voice of the AGENT only. If the agent has spoken in the prior context, describe that voice in detail (pitch/tone/pace/texture, e.g., \"low-pitched with slight rasp\"). If the agent has NOT yet spoken in the context, you MUST ignore the interlocutor\u2019s voice entirely and infer a new, plausible voice for the AGENT EXCLUSIVELY based on his visual appearance (age, gender, physique). (b) Emotional progression (e.g., \"reflective shifting to joyful tone\"). (c) Recording environment ONLY if it affects vocal quality (e.g., \"dry and direct, quiet room\"). - NEVER mention background sounds, other voices, or specific locations (e.g., \"gift shop\"). - Never mix or attribute the interlocutor\u2019s vocal traits to the agent.\nTEMPORAL & STYLE RULES: - Chronological flow: [Full appearance sentence] \u2192 [Pre-speech actions] \u2192 [<S>Speech Content<E>] \u2192 [Post-speech actions] \u2192 (repeat for multiple speech segments if any) \u2192 Audio line - Appearance sentence must start with appearance-based identifier (e.g., \"A woman in a blue dress...\"), then use pronouns (she/he/they). - Use PRESENT TENSE throughout (e.g., \"she wears\", \"he gestures\"). - Write in THIRD-PERSON, single continuous paragraph. NO bullet points, headings, or line breaks. - NEVER use: agent, assistant, interlocutor, respond, reply, listen, understand.\nOUTPUT EXAMPLES (CRITICAL FOR FORMATTING): EXAMPLE 1 (Natural pause before speech, no target provided): \"A man in a grey sweatshirt stands before dark red curtains. His eyes track the interlocutor for two seconds, meeting the camera with a steady gaze, <S>I have been considering your proposal all week.<E> straightening his shoulders as his thumb brushes his beard thoughtfully. Audio: Calm low-pitched male voice with steady rhythm and slight rasp, recorded in a quiet room.\"\nEXAMPLE 2 (Silent reaction only, no target provided): \"A woman in a dark velvet head covering sits in stark black and white lighting. Her gaze drifts downward for three seconds, eyelids lowering slowly. Fingers tap once on the armrest as her head tilts upward with a subtle smile, eyes crinkling at the corners while maintaining silence. Audio: SILENT.\"\nEXAMPLE 3 (Multiple speech segments with emotional progression, no target provided):\nEXAMPLE 4 (Target provided and speech break into multiple segments, [Target Speech]: \"I completely understand your frustration with the current situation, but we need to stay focused on the long-term goals of the project.\"): \"A woman in a silk blouse sits against a bright window. She nods slowly as the interlocutor finishes, then takes a deep breath. Meeting the camera with a steady gaze, <S>I completely understand your frustration with the current situation,<E> her expression softens. Pausing briefly to adjust her posture, she continues, <S>but we need to stay focused on the long-term goals of the project.<E> She remains attentive, her eyes reflecting deep thought while awaiting a reply. Audio: Articulate female voice with a calm, reassuring tone and steady pace, recorded in a quiet environment.\"\nEXAMPLE 5 (Target provided andcontains [Target State]: Listening): \"A man in a flannel shirt leans back against a brick wall. He tilts his head slightly to the right, eyes tracking the interlocutor\u2019s movements. His brow furrows momentarily in concentration, lips pressed together in a neutral line. He maintains a steady, receptive posture, nodding once in acknowledgement while remaining silent. Audio: SILENT.\"\nNEVER OUTPUT LIKE THIS (common failures): - \"...last words of agent\u2019s context speech. <S>Last words of context, then new content.<E>\" // Repeating context - \"The agent looks thoughtful after hearing the question.\"  // Banned words + internal state - \"<S>I agree<E> <S>But I don\u2019t think it\u2019s a good idea<E>\"  // Short speech broken down into multiple segments and not embedded in actions - \"Audio: Quiet room with distant traffic.\"  // Illegal background sound - \"A man speaks about his recovery.\"  // Invented topic (must derive from context) - \"She has a kind face.\"  // Abstract trait without visible cues",
+  "target_same_timestep_across_blocks": false,
+  "teacher_checkpoint": null,
+  "teacher_lora_checkpointing": {},
+  "text_encoder_id": "./pretrained_models/Wan2.2-TI2V-5B",
+  "text_encoder_path": "./pretrained_models/Wan2.2-TI2V-5B/models_t5_umt5-xxl-enc-bf16.pth",
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.3",
+  "use_ref_as_sink": true,
+  "use_ref_audio_concat": {
+    "enable": true,
+    "uncond_ratio": 0.0
+  },
+  "vae_downsample_f": 16,
+  "vae_downsample_t": 4,
+  "vae_id": "Lightricks/LTX-Video",
+  "vae_path": "./pretrained_models/Wan2.2-TI2V-5B/Wan2.2_VAE.pth",
+  "video_context_noise": 0.5,
+  "video_lora_dict": {
+    "adapter_name": "s2_ovi_vid_lora",
+    "enable": true,
+    "lora_dropout": 0.0,
+    "network_alpha": 128,
+    "rank": 128,
+    "target_modules": [
+      "q",
+      "k",
+      "v",
+      "o",
+      "ffn.0",
+      "ffn.2"
+    ]
+  },
+  "video_negative_prompt": "jitter, bad hands, blur, distortion",
+  "window": {}
+}

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/global_step5900/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4b9eeb1173b423c21868fdcd4c49ae2cf24dda45ee9f451c890e87c5e25c1e2
+size 967335749

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/global_step5900/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d3cdc1087c1c99f49e5e02e5679c7bf51bffa19041bc0c21b381a77ca5bcfe8
+size 967334917

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/global_step5900/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:284f45325943d0b8593e940196c0ac0e61b432bfe31a09840a7bcb7e98d1bb3c
+size 967335877

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/global_step5900/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:569a4c2b82ff9b3e38101b7042077f46cbec3012555d1956032b07c44f767c52
+size 967334917

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/global_step5900/bf16_zero_pp_rank_4_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ef19b7a8503e46df4a2ca9908345cf16f7d6a4a101fc598e84a161959df88cba
+size 967336069

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/global_step5900/bf16_zero_pp_rank_5_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d4c28f736d0ae1df193328784da0d7dc6f603b7e9f942041c5311f3cc645b45
+size 967335173

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/global_step5900/bf16_zero_pp_rank_6_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:612746816ccde0e0719d52a60519381512fe8c537a9a26e9e6512c7ec7cd78b5
+size 967336197

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/global_step5900/bf16_zero_pp_rank_7_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f4dea15618e604721b3d80624ddae92fd0fcd796531829b7508f95b963785e2
+size 967335237

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/global_step5900/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2cea3039335307499cf89c7705d5e513d8be26d596432c446eabef94c5fb5ad4
+size 3871280230

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/latest

@@ -0,0 +1 @@
+global_step5900

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d9a87cafa2c5d1e3bd60e1ba7ce57a4f228964facb92b20a93796d66c90cef5a
+size 3869833696

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cbef818631e190b671ab1352d4c9c520b1913ebe4e069333a552e4555230487e
+size 16325

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d59a4bcaf1a5cad5bdef123596828ae2c16102f9e6f86b56ae1a187118a3428
+size 16325

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bb9ec0e4e639232b6f53059d1fcaaee0516838431ec7de823d15cf71ce9c6a93
+size 16325

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ccc054f977cabd74ebf2395f556bd0af1aebe95951181c06e9e982114eca25ea
+size 16325

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/rng_state_4.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e767fe03beb54719316b70bc01ca7c9fd19f506a0fee5ebe67f842c07e3378bf
+size 16325

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/rng_state_5.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e3a2261885b31bd3f9ecce2a5c3b02db3dee5a085d32522ead75a775bad9a5ef
+size 16325

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/rng_state_6.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:962a54cfbbaae72f605fd41770260555c030920861e153367525c0ae032b3dfe
+size 16325

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/rng_state_7.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4d7b44fef562c9abcb996ec5fd595615405efcbf77537a887a7d48c7dbf849a1
+size 16325

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:308c4af63db2c978441c71cb60140c71a20f6a43c1e2b6f154968f3e865eef44
+size 1273

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/stage4_generator_optimizer.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed339a900b5f1ff7ba6f0d7bdf1f3327525b559c113ad6ac5d76512f17042fa1
+size 2580592297

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:314102a8d6a1a76859191599d91c549087e90c8cb0a8219096bf3ddeb91436a1
+size 8337

backup-0727/output/0224_stage4_self_rollout_dmd/backup/checkpoint-5900/checkpoint-5900/zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if ZERO_STAGE not in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info("Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info("Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)

backup-0727/output/0224_stage4_self_rollout_dmd/checkpoint-9900/checkpoint-9900/config.json

@@ -0,0 +1,356 @@
+{
+  "_gradient_checkpointing": true,
+  "ar_state_text_system_prompt": "You are a helpful assistant, capable of perceiving auditory and textual inputs, as well as generating text responses.",
+  "ar_target_debug": true,
+  "ar_target_do_sample": true,
+  "ar_target_max_new_tokens": 64,
+  "ar_target_temperature": 0.9,
+  "ar_target_top_k": 50,
+  "ar_target_top_p": 0.9,
+  "ar_target_use_v2": true,
+  "architectures": [
+    "Stage4DMDSelfRolloutMetaQuery"
+  ],
+  "attn_implementation": "auto",
+  "audio_context_noise": 0.5,
+  "audio_lora_dict": {
+    "adapter_name": "s2_ovi_audio_lora",
+    "enable": true,
+    "lora_dropout": 0.0,
+    "network_alpha": 128,
+    "rank": 128,
+    "target_modules": [
+      "q",
+      "k",
+      "v",
+      "o",
+      "ffn.0",
+      "ffn.2"
+    ]
+  },
+  "audio_loss_weight": 0.5,
+  "audio_negative_prompt": "robotic, muffled, echo, distorted",
+  "audio_vae_path": "./pretrained_models/MMAudio/ext_weights/",
+  "block_causal": {
+    "a2a_lookahead": -1,
+    "attention_backend": "sdpa",
+    "audio_kv_cache_size": 4,
+    "enable": true,
+    "frame_per_block": 3,
+    "teacher_forcing": true,
+    "v2a_lookahead": 1,
+    "video_kv_cache_size": 4
+  },
+  "bsmntw_weighting": true,
+  "connector_num_hidden_layers": 8,
+  "diffusion_forcing": {
+    "enable": false,
+    "prefix_only": false,
+    "prob": 0.9,
+    "start_step": 0,
+    "temp_chunk": 3
+  },
+  "diffusion_model_id": "Lightricks/LTX-Video",
+  "diffusion_model_path": "./pretrained_models/Ovi/model_960x960_10s.safetensors",
+  "distillation": {},
+  "dmd": {
+    "audio_loss_weight": 0.5,
+    "critic_audio_loss_weight": 0.5,
+    "critic_video_loss_weight": 1.0,
+    "denoising_step_lists": {
+      "1": [
+        1000
+      ],
+      "2": [
+        1000,
+        500
+      ],
+      "3": [
+        1000,
+        750,
+        500
+      ],
+      "4": [
+        1000,
+        750,
+        500,
+        250
+      ]
+    },
+    "denoising_step_lists_first_chunk": {
+      "1": [
+        1000,
+        750,
+        500,
+        250
+      ],
+      "2": [
+        1000,
+        750,
+        500,
+        250
+      ],
+      "3": [
+        1000,
+        750,
+        500,
+        250
+      ],
+      "4": [
+        1000,
+        750,
+        500,
+        250
+      ]
+    },
+    "detach_fusion_audio": false,
+    "detach_fusion_video": false,
+    "detach_mllm_mq": true,
+    "ema_start_step": 200,
+    "ema_weight": 0.99,
+    "fake_guidance_scale": 0.0,
+    "fake_score_learning_rate": 5e-06,
+    "fake_score_update_interval": 5,
+    "flowmatch_shift": 5.0,
+    "gan_audio_start_step": 0,
+    "gan_critic_checkpoint_blocks": 30,
+    "gan_feature_depths": [
+      0.1,
+      0.5,
+      0.9
+    ],
+    "gan_loss_weight_audio": 0.0,
+    "gan_loss_weight_video": 0.0,
+    "gan_t_max": 1000,
+    "gan_t_min": 0,
+    "gan_video_start_step": 0,
+    "generator_adam_beta1": 0.0,
+    "generator_audio_loss_weight": 0.5,
+    "generator_gt_audio_listening_weight_multiplier": 1.0,
+    "generator_gt_audio_loss_weight": 0.05,
+    "generator_gt_video_loss_weight": 0.0,
+    "generator_video_loss_weight": 1.0,
+    "gt_score_regularize_start_step": 93000,
+    "gt_score_regularize_weight": 0.5,
+    "lookahead_weight": 2.0,
+    "lookahead_weight_position": 1,
+    "min_score_timestep": 0,
+    "normalize_gradient": true,
+    "num_inference_steps": 2,
+    "num_train_timestep": 1000,
+    "plan_commit_consistency_enable": false,
+    "plan_commit_gt_weight": 1.0,
+    "plan_commit_loss_weight": 3.0,
+    "purify_draft_static_margin": 4.0,
+    "purify_draft_static_mask_enable": false,
+    "purify_draft_static_quantile": 0.2,
+    "purify_draft_weight": 0.0,
+    "purify_start_step": 0,
+    "purify_t_max": 100,
+    "purify_t_min": 0,
+    "purify_weight": 0.0,
+    "real_audio_scale": 4.5,
+    "real_guidance_scale": 3.5,
+    "real_score_gt_x0_ratio": 0.5,
+    "real_score_gt_x0_start_step": 36000,
+    "real_slg_layer": 11,
+    "score_flowmatch_shift": 5.0,
+    "trainer": "dmd_self_rollout",
+    "ts_schedule": false,
+    "ts_schedule_max": false,
+    "validation_score_timestep": 250,
+    "video_loss_weight": 1.0,
+    "warp_denoising_step": true
+  },
+  "fake_score_learning_rate": 5e-06,
+  "fake_score_lora_checkpointing": {
+    "audio_adapter_name": "s4_fake_ovi_audio_lora",
+    "block_causal": {
+      "enable": false
+    },
+    "merge_from_checkpoint": "./output/0217_s1_qwen2p5omni7b_ovi1p1_mq_nonverbal_context_and_T5_response_talking80_vision_audio_loss_0p5_ICL_voice_prefix_sink_loss",
+    "merge_load_non_lora": true,
+    "resume_from_checkpoint": null,
+    "resume_gan_head": false,
+    "resume_lora_only": false,
+    "video_adapter_name": "s4_fake_ovi_vid_lora"
+  },
+  "flowmatch_shift": 5.0,
+  "gen_fps": 25.0,
+  "generator_lora_checkpointing": {
+    "audio_adapter_name": "s2_ovi_audio_lora",
+    "merge_from_checkpoint": "./output/0217_s1_qwen2p5omni7b_ovi1p1_mq_nonverbal_context_and_T5_response_talking80_vision_audio_loss_0p5_ICL_voice_prefix_sink_loss",
+    "merge_load_non_lora": true,
+    "prefer_ema": true,
+    "resume_from_checkpoint": "./output/0222_stage3_causal_cd_frame3",
+    "resume_lora_only": true,
+    "video_adapter_name": "s2_ovi_vid_lora"
+  },
+  "has_context_prompt": {
+    "enable": true,
+    "gt_t5_transcript_ratio": 1.0,
+    "no_mq_context_uncond_ratio": 0.0,
+    "skip_mllm_encode_ratio": 0.3
+  },
+  "in_channels": 16,
+  "infer": false,
+  "latent_spatial_size": 28,
+  "latent_temporal_size": 25,
+  "listening_loss_weight": 1.0,
+  "listening_prompt": "not speaking, closed mouth, listening, silent",
+  "local_files_only": true,
+  "lora_checkpointing": {
+    "audio_adapter_name": "s2_ovi_audio_lora",
+    "merge_from_checkpoint": "./output/0217_s1_qwen2p5omni7b_ovi1p1_mq_nonverbal_context_and_T5_response_talking80_vision_audio_loss_0p5_ICL_voice_prefix_sink_loss",
+    "merge_load_non_lora": true,
+    "prefer_ema": true,
+    "resume_from_checkpoint": "./output/0222_stage3_causal_cd_frame3",
+    "resume_lora_only": true,
+    "video_adapter_name": "s2_ovi_vid_lora"
+  },
+  "lora_dict": {},
+  "max_input_text_tokens": 256,
+  "mllm_ce_only": false,
+  "mllm_id": "./pretrained_models/Qwen2.5-Omni-7B",
+  "mllm_local_path": null,
+  "mllm_lora_dict": {
+    "alpha": 64,
+    "dropout": 0.0,
+    "enable": false,
+    "rank": 32,
+    "target_modules": [
+      "q_proj",
+      "k_proj",
+      "v_proj",
+      "o_proj",
+      "gate_proj",
+      "up_proj",
+      "down_proj"
+    ]
+  },
+  "model_type": "stage4_dmd_self_rollout_metaquery",
+  "modules_to_freeze": [
+    "vae_model_video",
+    "vae_model_audio",
+    "text_model",
+    "fusion_model",
+    "mllm.mllm_backbone",
+    "mllm.connector"
+  ],
+  "modules_to_unfreeze": [],
+  "negative_text_prompt": "",
+  "noise_scheduler_id": "Lightricks/LTX-Video",
+  "num_metaqueries": 128,
+  "ovi_gradient_checkpointing_block_count": null,
+  "prev_temp_chunk": 1,
+  "real_score_lora_checkpointing": {
+    "audio_adapter_name": "s4_real_ovi_audio_lora",
+    "block_causal": {
+      "enable": false
+    },
+    "merge_from_checkpoint": "./output/0217_s1_qwen2p5omni7b_ovi1p1_mq_nonverbal_context_and_T5_response_talking80_vision_audio_loss_0p5_ICL_voice_prefix_sink_loss",
+    "merge_load_non_lora": true,
+    "resume_from_checkpoint": null,
+    "resume_lora_only": false,
+    "video_adapter_name": "s4_real_ovi_vid_lora"
+  },
+  "rollout": {
+    "agent_context_source": "gt",
+    "aud_kv_gt_ratio": 0.0,
+    "audio_exit_replay_grad": false,
+    "audio_kv_cache_size": 4,
+    "audio_kv_grad_blocks": 0,
+    "context_noise": 0,
+    "decode_context_min_video_latents": 12,
+    "enable_ode_step": false,
+    "exit_index_sync": "broadcast",
+    "extra_kv_clean_forward": false,
+    "kv_cache_size": 4,
+    "kv_shift_step": 0,
+    "last_step_only": false,
+    "listening_output_fade_samples": 1536,
+    "listening_silence_enable": false,
+    "num_windows_schedule": [
+      [
+        0,
+        [
+          [
+            1
+          ],
+          [
+            1
+          ]
+        ]
+      ],
+      [
+        500,
+        [
+          [
+            1,
+            2,
+            3
+          ],
+          [
+            0.4,
+            0.4,
+            0.2
+          ]
+        ]
+      ]
+    ],
+    "real_score_bridge_use_gt": false,
+    "same_step_across_blocks": true,
+    "score_bridge_vid_latents": 5,
+    "share_video_noise_weight": 0,
+    "use_kv_cache": true,
+    "user_context_source": "gt",
+    "validation_decode": true,
+    "validation_max_samples": 1,
+    "validation_num_windows": 3,
+    "validation_seed": 920,
+    "vid_kv_gt_ratio": 0.0,
+    "video_kv_cache_size": 4,
+    "video_latents_per_window": 24,
+    "window_duration_sec": 3.84
+  },
+  "scheduler_id": "Lightricks/LTX-Video",
+  "stage": 4,
+  "stage_name": "dmd_self_rollout",
+  "student_checkpoint": null,
+  "student_lora_checkpointing": {},
+  "system_prompt": "You are a speech-video clip narrator for a text-to-audio-video (T2AV) model.\nThe input includes audio-visual dialogue between the Interlocutor and the Agent. The dialogue context is provided ONLY to infer what occurs next and must NOT be described, referenced, or summarized explicitly.\nYour task is to write a narration describing the NEXT recorded speech and video clip that shows ONLY the AGENT. Follow these rules precisely:\nVISUAL RULES: - Provide a HIGH-DETAIL appearance sentence, covering: (a) Gender + precise clothing (material/color/style, e.g., \"velvet head covering with ruffled accent\"), (b) Distinctive features (eyeliner/beard/accessories), (c) Scene composition (framing/lighting/color tone/backdrop, e.g., \"stark black and white lighting against plain curtains\"). - Describe ONLY visible physical actions in strict chronological order: (a) Facial muscles (e.g., \"lips curve upward forming a subtle smile\"), (b) Eye movements (e.g., \"gaze drifts downward for two seconds\"), (c) Posture/gestures (e.g., \"thumb brushes beard thoughtfully\"). - Emotions MUST derive from observable cues: CORRECT: \"eyes crinkling at the corners in a warm smile\" (visible) WRONG: \"she feels happy\" (internal) or \"looks thoughtful\" (abstract).\nSPEECH RULES: - Determine speech occurrence based on PROVIDED TARGETS or NATURAL CONVERSATION FLOW: - IF TARGETS ARE PROVIDED: (a) If \"[Target State]: Listening\" is provided in the context, describe silent actions ONLY without any speech content and do not use <S><E> tags. (b) If \"[Target Speech]: [Text]\" is provided, use the provided [Text] exactly as full speech content. - IF TARGETS ARE NOT PROVIDED: (a) If dialogue context implies continued listening or non-verbal reaction, describe silent actions ONLY without any speech content and do not use <S><E> tags. (b) If dialogue context implies the agent WILL SPEAK NEXT (even after brief listening actions), generate speech. (c) When speech needs to be generated: (1) Infer the speech content EXCLUSIVELY from dialogue context. (2) The speech content must be a RESPONSE that move forward the conversation, but word overlap with the end of the agent\u2019s speech context is STRICTLY FORBIDDEN. - GENERAL SPEECH STYLE (Only for speech existing in the narration): (b) Wrap each segment in <S> and <E> (speech contents are broken down into multiple segments when there are too many words for one segment) in the MIDDLE of the narration. (c) For each speech segment, embed it NATURALLY within action descriptions BEFORE and AFTER speech: CORRECT: \"meeting the camera, <S>I have considered your proposal<E> he brushes his thumb against his beard.\" WRONG: separate speech/action blocks.\nAUDIO RULES: - Add \"Audio: [description]\" at the VERY END of the paragraph, separated by a single space. - If no speech occurs (means no <S><E> tags exist), use \"SILENT\" for the audio description. - Otherwise, describe ONLY the agent\u2019s audio characteristics: (a) Voice traits: Describe the voice of the AGENT only. If the agent has spoken in the prior context, describe that voice in detail (pitch/tone/pace/texture, e.g., \"low-pitched with slight rasp\"). If the agent has NOT yet spoken in the context, you MUST ignore the interlocutor\u2019s voice entirely and infer a new, plausible voice for the AGENT EXCLUSIVELY based on his visual appearance (age, gender, physique). (b) Emotional progression (e.g., \"reflective shifting to joyful tone\"). (c) Recording environment ONLY if it affects vocal quality (e.g., \"dry and direct, quiet room\"). - NEVER mention background sounds, other voices, or specific locations (e.g., \"gift shop\"). - Never mix or attribute the interlocutor\u2019s vocal traits to the agent.\nTEMPORAL & STYLE RULES: - Chronological flow: [Full appearance sentence] \u2192 [Pre-speech actions] \u2192 [<S>Speech Content<E>] \u2192 [Post-speech actions] \u2192 (repeat for multiple speech segments if any) \u2192 Audio line - Appearance sentence must start with appearance-based identifier (e.g., \"A woman in a blue dress...\"), then use pronouns (she/he/they). - Use PRESENT TENSE throughout (e.g., \"she wears\", \"he gestures\"). - Write in THIRD-PERSON, single continuous paragraph. NO bullet points, headings, or line breaks. - NEVER use: agent, assistant, interlocutor, respond, reply, listen, understand.\nOUTPUT EXAMPLES (CRITICAL FOR FORMATTING): EXAMPLE 1 (Natural pause before speech, no target provided): \"A man in a grey sweatshirt stands before dark red curtains. His eyes track the interlocutor for two seconds, meeting the camera with a steady gaze, <S>I have been considering your proposal all week.<E> straightening his shoulders as his thumb brushes his beard thoughtfully. Audio: Calm low-pitched male voice with steady rhythm and slight rasp, recorded in a quiet room.\"\nEXAMPLE 2 (Silent reaction only, no target provided): \"A woman in a dark velvet head covering sits in stark black and white lighting. Her gaze drifts downward for three seconds, eyelids lowering slowly. Fingers tap once on the armrest as her head tilts upward with a subtle smile, eyes crinkling at the corners while maintaining silence. Audio: SILENT.\"\nEXAMPLE 3 (Multiple speech segments with emotional progression, no target provided):\nEXAMPLE 4 (Target provided and speech break into multiple segments, [Target Speech]: \"I completely understand your frustration with the current situation, but we need to stay focused on the long-term goals of the project.\"): \"A woman in a silk blouse sits against a bright window. She nods slowly as the interlocutor finishes, then takes a deep breath. Meeting the camera with a steady gaze, <S>I completely understand your frustration with the current situation,<E> her expression softens. Pausing briefly to adjust her posture, she continues, <S>but we need to stay focused on the long-term goals of the project.<E> She remains attentive, her eyes reflecting deep thought while awaiting a reply. Audio: Articulate female voice with a calm, reassuring tone and steady pace, recorded in a quiet environment.\"\nEXAMPLE 5 (Target provided andcontains [Target State]: Listening): \"A man in a flannel shirt leans back against a brick wall. He tilts his head slightly to the right, eyes tracking the interlocutor\u2019s movements. His brow furrows momentarily in concentration, lips pressed together in a neutral line. He maintains a steady, receptive posture, nodding once in acknowledgement while remaining silent. Audio: SILENT.\"\nNEVER OUTPUT LIKE THIS (common failures): - \"...last words of agent\u2019s context speech. <S>Last words of context, then new content.<E>\" // Repeating context - \"The agent looks thoughtful after hearing the question.\"  // Banned words + internal state - \"<S>I agree<E> <S>But I don\u2019t think it\u2019s a good idea<E>\"  // Short speech broken down into multiple segments and not embedded in actions - \"Audio: Quiet room with distant traffic.\"  // Illegal background sound - \"A man speaks about his recovery.\"  // Invented topic (must derive from context) - \"She has a kind face.\"  // Abstract trait without visible cues",
+  "target_same_timestep_across_blocks": false,
+  "teacher_checkpoint": null,
+  "teacher_lora_checkpointing": {},
+  "text_encoder_id": "./pretrained_models/Wan2.2-TI2V-5B",
+  "text_encoder_path": "./pretrained_models/Wan2.2-TI2V-5B/models_t5_umt5-xxl-enc-bf16.pth",
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.3",
+  "use_ref_as_sink": true,
+  "use_ref_audio_concat": {
+    "enable": true,
+    "uncond_ratio": 0.0
+  },
+  "vae_downsample_f": 16,
+  "vae_downsample_t": 4,
+  "vae_id": "Lightricks/LTX-Video",
+  "vae_path": "./pretrained_models/Wan2.2-TI2V-5B/Wan2.2_VAE.pth",
+  "video_context_noise": 0.5,
+  "video_lora_dict": {
+    "adapter_name": "s2_ovi_vid_lora",
+    "enable": true,
+    "lora_dropout": 0.0,
+    "network_alpha": 128,
+    "rank": 128,
+    "target_modules": [
+      "q",
+      "k",
+      "v",
+      "o",
+      "ffn.0",
+      "ffn.2"
+    ]
+  },
+  "video_negative_prompt": "jitter, bad hands, blur, distortion",
+  "window": {}
+}

backup-0727/output/0224_stage4_self_rollout_dmd/checkpoint-9900/checkpoint-9900/latest

@@ -0,0 +1 @@
+global_step9900

backup-0727/output/0224_stage4_self_rollout_dmd/checkpoint-9900/checkpoint-9900/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d00a3fbbe1b8c8a8aa4db85cb0ccea9ad71b526e4d65a19de59ce0a68715c432
+size 3869833696

backup-0727/output/0224_stage4_self_rollout_dmd/checkpoint-9900/checkpoint-9900/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e94a3ca0cf2fc17e7c7eb0260dfa3fcbcb873636fccff771e831d7dbdea1d82d
+size 16389

backup-0727/output/0224_stage4_self_rollout_dmd/checkpoint-9900/checkpoint-9900/rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:84d2177320bea72a95af63721ceed6f47eb3a5b023d4e7ae950580ae956390c0
+size 16389

backup-0727/output/0224_stage4_self_rollout_dmd/checkpoint-9900/checkpoint-9900/rng_state_4.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a9293c15981aad85745ecbdd8d17bcac42a886923d7f7cfec9f7501a7f53108
+size 16389

backup-0727/output/0224_stage4_self_rollout_dmd/checkpoint-9900/checkpoint-9900/rng_state_5.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:87d885dd44166e228b1316585401bcb1c26c6eb43963ec6b187527ada5cd6dc9
+size 16389

backup-0727/output/0224_stage4_self_rollout_dmd/checkpoint-9900/checkpoint-9900/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:314102a8d6a1a76859191599d91c549087e90c8cb0a8219096bf3ddeb91436a1
+size 8337

backup-0727/output/0224_stage4_self_rollout_dmd/checkpoint-9900/checkpoint-9900/zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if ZERO_STAGE not in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info("Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info("Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)

backup-260717/0217_s1_qwen2p5omni7b_ovi1p1_mq_nonverbal_context_and_T5_response_talking80_vision_audio_loss_0p5_ICL_voice_prefix_sink_loss-checkpoint-29400/config.json

@@ -0,0 +1,154 @@
+{
+  "_gradient_checkpointing": true,
+  "ar_state_text_system_prompt": "You are a helpful assistant, capable of perceiving auditory and textual inputs, as well as generating text responses.",
+  "ar_target_debug": true,
+  "ar_target_do_sample": true,
+  "ar_target_max_new_tokens": 64,
+  "ar_target_temperature": 0.9,
+  "ar_target_top_k": 50,
+  "ar_target_top_p": 0.9,
+  "ar_target_use_v2": true,
+  "architectures": [
+    "MetaQuery"
+  ],
+  "attn_implementation": "auto",
+  "audio_context_noise": 0.0,
+  "audio_lora_dict": {
+    "adapter_name": "s1_ovi_audio_lora",
+    "enable": true,
+    "lora_dropout": 0.0,
+    "network_alpha": 128,
+    "rank": 128,
+    "target_modules": [
+      "q",
+      "k",
+      "v",
+      "o",
+      "ffn.0",
+      "ffn.2"
+    ]
+  },
+  "audio_loss_weight": 0.5,
+  "audio_negative_prompt": "robotic, muffled, echo, distorted",
+  "audio_vae_path": "./pretrained_models/MMAudio/ext_weights/",
+  "block_causal": {
+    "enable": false,
+    "frame_per_block": 1
+  },
+  "bsmntw_weighting": true,
+  "connector_num_hidden_layers": 8,
+  "diffusion_forcing": {
+    "enable": true,
+    "prefix_only": true,
+    "prob": 0.65,
+    "start_step": 4000,
+    "temp_chunk": 3
+  },
+  "diffusion_model_id": "Lightricks/LTX-Video",
+  "diffusion_model_path": "./pretrained_models/Ovi/model_960x960_10s.safetensors",
+  "distillation": {},
+  "dmd": {},
+  "fake_score_learning_rate": null,
+  "fake_score_lora_checkpointing": {},
+  "flowmatch_shift": 5.0,
+  "gen_fps": 25.0,
+  "generator_lora_checkpointing": {},
+  "has_context_prompt": {
+    "enable": true,
+    "gt_t5_transcript_ratio": 1.0,
+    "no_mq_context_uncond_ratio": 0.25,
+    "skip_mllm_encode_ratio": 0.4
+  },
+  "in_channels": 16,
+  "infer": false,
+  "latent_spatial_size": 28,
+  "latent_temporal_size": 29,
+  "listening_loss_weight": 1.0,
+  "listening_prompt": "not speaking, closed mouth, listening, silent",
+  "local_files_only": true,
+  "lora_checkpointing": {
+    "merge_from_checkpoint": null,
+    "merge_load_non_lora": true,
+    "resume_from_checkpoint": null,
+    "resume_lora_only": false
+  },
+  "lora_dict": {},
+  "max_input_text_tokens": 256,
+  "mllm_ce_only": false,
+  "mllm_id": "./pretrained_models/Qwen2.5-Omni-7B",
+  "mllm_local_path": null,
+  "mllm_lora_dict": {
+    "alpha": 64,
+    "dropout": 0.0,
+    "enable": false,
+    "rank": 32,
+    "target_modules": [
+      "q_proj",
+      "k_proj",
+      "v_proj",
+      "o_proj",
+      "gate_proj",
+      "up_proj",
+      "down_proj"
+    ]
+  },
+  "model_type": "metaquery",
+  "modules_to_freeze": [
+    "vae_model_video",
+    "vae_model_audio",
+    "text_model",
+    "fusion_model",
+    "mllm.mllm_backbone"
+  ],
+  "modules_to_unfreeze": [
+    "mllm.connector",
+    "mllm.mllm_backbone.model.embed_tokens"
+  ],
+  "negative_text_prompt": "jitter, bad hands, blur, distortion, robotic, muffled, echo, distorted",
+  "noise_scheduler_id": "Lightricks/LTX-Video",
+  "num_metaqueries": 128,
+  "ovi_gradient_checkpointing_block_count": 25,
+  "prev_temp_chunk": 5,
+  "real_score_lora_checkpointing": {},
+  "rollout": {},
+  "scheduler_id": "Lightricks/LTX-Video",
+  "stage": 1,
+  "stage_name": null,
+  "student_checkpoint": null,
+  "student_lora_checkpointing": {},
+  "system_prompt": "You are a speech-video clip narrator for a text-to-audio-video (T2AV) model.\nThe input includes audio-visual dialogue between the Interlocutor and the Agent. The dialogue context is provided ONLY to infer what occurs next and must NOT be described, referenced, or summarized explicitly.\nYour task is to write a narration describing the NEXT recorded speech and video clip that shows ONLY the AGENT. Follow these rules precisely:\nVISUAL RULES: - Provide a HIGH-DETAIL appearance sentence, covering: (a) Gender + precise clothing (material/color/style, e.g., \"velvet head covering with ruffled accent\"), (b) Distinctive features (eyeliner/beard/accessories), (c) Scene composition (framing/lighting/color tone/backdrop, e.g., \"stark black and white lighting against plain curtains\"). - Describe ONLY visible physical actions in strict chronological order: (a) Facial muscles (e.g., \"lips curve upward forming a subtle smile\"), (b) Eye movements (e.g., \"gaze drifts downward for two seconds\"), (c) Posture/gestures (e.g., \"thumb brushes beard thoughtfully\"). - Emotions MUST derive from observable cues: CORRECT: \"eyes crinkling at the corners in a warm smile\" (visible) WRONG: \"she feels happy\" (internal) or \"looks thoughtful\" (abstract).\nSPEECH RULES: - Determine speech occurrence based on PROVIDED TARGETS or NATURAL CONVERSATION FLOW: - IF TARGETS ARE PROVIDED: (a) If \"[Target State]: Listening\" is provided in the context, describe silent actions ONLY without any speech content and do not use <S><E> tags. (b) If \"[Target Speech]: [Text]\" is provided, use the provided [Text] exactly as full speech content. - IF TARGETS ARE NOT PROVIDED: (a) If dialogue context implies continued listening or non-verbal reaction, describe silent actions ONLY without any speech content and do not use <S><E> tags. (b) If dialogue context implies the agent WILL SPEAK NEXT (even after brief listening actions), generate speech. (c) When speech needs to be generated: (1) Infer the speech content EXCLUSIVELY from dialogue context. (2) The speech content must be a RESPONSE that move forward the conversation, but word overlap with the end of the agent\u2019s speech context is STRICTLY FORBIDDEN. - GENERAL SPEECH STYLE (Only for speech existing in the narration): (b) Wrap each segment in <S> and <E> (speech contents are broken down into multiple segments when there are too many words for one segment) in the MIDDLE of the narration. (c) For each speech segment, embed it NATURALLY within action descriptions BEFORE and AFTER speech: CORRECT: \"meeting the camera, <S>I have considered your proposal<E> he brushes his thumb against his beard.\" WRONG: separate speech/action blocks.\nAUDIO RULES: - Add \"Audio: [description]\" at the VERY END of the paragraph, separated by a single space. - If no speech occurs (means no <S><E> tags exist), use \"SILENT\" for the audio description. - Otherwise, describe ONLY the agent\u2019s audio characteristics: (a) Voice traits: Describe the voice of the AGENT only. If the agent has spoken in the prior context, describe that voice in detail (pitch/tone/pace/texture, e.g., \"low-pitched with slight rasp\"). If the agent has NOT yet spoken in the context, you MUST ignore the interlocutor\u2019s voice entirely and infer a new, plausible voice for the AGENT EXCLUSIVELY based on his visual appearance (age, gender, physique). (b) Emotional progression (e.g., \"reflective shifting to joyful tone\"). (c) Recording environment ONLY if it affects vocal quality (e.g., \"dry and direct, quiet room\"). - NEVER mention background sounds, other voices, or specific locations (e.g., \"gift shop\"). - Never mix or attribute the interlocutor\u2019s vocal traits to the agent.\nTEMPORAL & STYLE RULES: - Chronological flow: [Full appearance sentence] \u2192 [Pre-speech actions] \u2192 [<S>Speech Content<E>] \u2192 [Post-speech actions] \u2192 (repeat for multiple speech segments if any) \u2192 Audio line - Appearance sentence must start with appearance-based identifier (e.g., \"A woman in a blue dress...\"), then use pronouns (she/he/they). - Use PRESENT TENSE throughout (e.g., \"she wears\", \"he gestures\"). - Write in THIRD-PERSON, single continuous paragraph. NO bullet points, headings, or line breaks. - NEVER use: agent, assistant, interlocutor, respond, reply, listen, understand.\nOUTPUT EXAMPLES (CRITICAL FOR FORMATTING): EXAMPLE 1 (Natural pause before speech, no target provided): \"A man in a grey sweatshirt stands before dark red curtains. His eyes track the interlocutor for two seconds, meeting the camera with a steady gaze, <S>I have been considering your proposal all week.<E> straightening his shoulders as his thumb brushes his beard thoughtfully. Audio: Calm low-pitched male voice with steady rhythm and slight rasp, recorded in a quiet room.\"\nEXAMPLE 2 (Silent reaction only, no target provided): \"A woman in a dark velvet head covering sits in stark black and white lighting. Her gaze drifts downward for three seconds, eyelids lowering slowly. Fingers tap once on the armrest as her head tilts upward with a subtle smile, eyes crinkling at the corners while maintaining silence. Audio: SILENT.\"\nEXAMPLE 3 (Multiple speech segments with emotional progression, no target provided):\nEXAMPLE 4 (Target provided and speech break into multiple segments, [Target Speech]: \"I completely understand your frustration with the current situation, but we need to stay focused on the long-term goals of the project.\"): \"A woman in a silk blouse sits against a bright window. She nods slowly as the interlocutor finishes, then takes a deep breath. Meeting the camera with a steady gaze, <S>I completely understand your frustration with the current situation,<E> her expression softens. Pausing briefly to adjust her posture, she continues, <S>but we need to stay focused on the long-term goals of the project.<E> She remains attentive, her eyes reflecting deep thought while awaiting a reply. Audio: Articulate female voice with a calm, reassuring tone and steady pace, recorded in a quiet environment.\"\nEXAMPLE 5 (Target provided andcontains [Target State]: Listening): \"A man in a flannel shirt leans back against a brick wall. He tilts his head slightly to the right, eyes tracking the interlocutor\u2019s movements. His brow furrows momentarily in concentration, lips pressed together in a neutral line. He maintains a steady, receptive posture, nodding once in acknowledgement while remaining silent. Audio: SILENT.\"\nNEVER OUTPUT LIKE THIS (common failures): - \"...last words of agent\u2019s context speech. <S>Last words of context, then new content.<E>\" // Repeating context - \"The agent looks thoughtful after hearing the question.\"  // Banned words + internal state - \"<S>I agree<E> <S>But I don\u2019t think it\u2019s a good idea<E>\"  // Short speech broken down into multiple segments and not embedded in actions - \"Audio: Quiet room with distant traffic.\"  // Illegal background sound - \"A man speaks about his recovery.\"  // Invented topic (must derive from context) - \"She has a kind face.\"  // Abstract trait without visible cues",
+  "target_same_timestep_across_blocks": false,
+  "teacher_checkpoint": null,
+  "teacher_lora_checkpointing": {},
+  "text_encoder_id": "./pretrained_models/Wan2.2-TI2V-5B",
+  "text_encoder_path": "./pretrained_models/Wan2.2-TI2V-5B/models_t5_umt5-xxl-enc-bf16.pth",
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.3",
+  "use_ref_as_sink": true,
+  "use_ref_audio_concat": {
+    "enable": true,
+    "uncond_ratio": 0.1
+  },
+  "vae_downsample_f": 16,
+  "vae_downsample_t": 4,
+  "vae_id": "Lightricks/LTX-Video",
+  "vae_path": "./pretrained_models/Wan2.2-TI2V-5B/Wan2.2_VAE.pth",
+  "video_context_noise": 0.0,
+  "video_lora_dict": {
+    "adapter_name": "s1_ovi_vid_lora",
+    "enable": true,
+    "lora_dropout": 0.0,
+    "network_alpha": 128,
+    "rank": 128,
+    "target_modules": [
+      "q",
+      "k",
+      "v",
+      "o",
+      "ffn.0",
+      "ffn.2"
+    ]
+  },
+  "video_negative_prompt": "jitter, bad hands, blur, distortion",
+  "window": {}
+}

backup-260717/0217_s1_qwen2p5omni7b_ovi1p1_mq_nonverbal_context_and_T5_response_talking80_vision_audio_loss_0p5_ICL_voice_prefix_sink_loss-checkpoint-29400/latest

@@ -0,0 +1 @@
+global_step29400

backup-260717/0217_s1_qwen2p5omni7b_ovi1p1_mq_nonverbal_context_and_T5_response_talking80_vision_audio_loss_0p5_ICL_voice_prefix_sink_loss-checkpoint-29400/zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if ZERO_STAGE not in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info("Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info("Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)

backup-260717/0220_stage2_ar_tf_frame3_context_noise_0p5_aud_lookahead_a2a_all_v2a_1-checkpoint-8800/config.json

@@ -0,0 +1,158 @@
+{
+  "_gradient_checkpointing": true,
+  "ar_state_text_system_prompt": "You are a helpful assistant, capable of perceiving auditory and textual inputs, as well as generating text responses.",
+  "ar_target_debug": true,
+  "ar_target_do_sample": true,
+  "ar_target_max_new_tokens": 64,
+  "ar_target_temperature": 0.9,
+  "ar_target_top_k": 50,
+  "ar_target_top_p": 0.9,
+  "ar_target_use_v2": true,
+  "architectures": [
+    "Stage2MetaQuery"
+  ],
+  "attn_implementation": "auto",
+  "audio_context_noise": 0.5,
+  "audio_lora_dict": {
+    "adapter_name": "s2_ovi_audio_lora",
+    "enable": true,
+    "lora_dropout": 0.0,
+    "network_alpha": 128,
+    "rank": 128,
+    "target_modules": [
+      "q",
+      "k",
+      "v",
+      "o",
+      "ffn.0",
+      "ffn.2"
+    ]
+  },
+  "audio_loss_weight": 0.5,
+  "audio_negative_prompt": "robotic, muffled, echo, distorted",
+  "audio_vae_path": "./pretrained_models/MMAudio/ext_weights/",
+  "block_causal": {
+    "a2a_lookahead": -1,
+    "attention_backend": "auto",
+    "audio_kv_cache_size": 4,
+    "enable": true,
+    "frame_per_block": 3,
+    "teacher_forcing": true,
+    "v2a_lookahead": 1,
+    "video_kv_cache_size": 4
+  },
+  "bsmntw_weighting": true,
+  "connector_num_hidden_layers": 8,
+  "diffusion_forcing": {
+    "enable": false,
+    "prefix_only": false,
+    "prob": 0.9,
+    "start_step": 0,
+    "temp_chunk": 3
+  },
+  "diffusion_model_id": "Lightricks/LTX-Video",
+  "diffusion_model_path": "./pretrained_models/Ovi/model_960x960_10s.safetensors",
+  "distillation": {},
+  "dmd": {},
+  "fake_score_learning_rate": null,
+  "fake_score_lora_checkpointing": {},
+  "flowmatch_shift": 5.0,
+  "gen_fps": 25.0,
+  "generator_lora_checkpointing": {},
+  "has_context_prompt": {
+    "enable": true,
+    "gt_t5_transcript_ratio": 1.0,
+    "no_mq_context_uncond_ratio": 0.25,
+    "skip_mllm_encode_ratio": 0.4
+  },
+  "in_channels": 16,
+  "infer": false,
+  "latent_spatial_size": 28,
+  "latent_temporal_size": 25,
+  "listening_loss_weight": 1.0,
+  "listening_prompt": "not speaking, closed mouth, listening, silent",
+  "local_files_only": true,
+  "lora_checkpointing": {
+    "merge_from_checkpoint": "./output/0217_s1_qwen2p5omni7b_ovi1p1_mq_nonverbal_context_and_T5_response_talking80_vision_audio_loss_0p5_ICL_voice_prefix_sink_loss",
+    "merge_load_non_lora": true,
+    "resume_from_checkpoint": null,
+    "resume_lora_only": false
+  },
+  "lora_dict": {},
+  "max_input_text_tokens": 256,
+  "mllm_ce_only": false,
+  "mllm_id": "./pretrained_models/Qwen2.5-Omni-7B",
+  "mllm_local_path": null,
+  "mllm_lora_dict": {
+    "alpha": 64,
+    "dropout": 0.0,
+    "enable": false,
+    "rank": 32,
+    "target_modules": [
+      "q_proj",
+      "k_proj",
+      "v_proj",
+      "o_proj",
+      "gate_proj",
+      "up_proj",
+      "down_proj"
+    ]
+  },
+  "model_type": "stage2_metaquery",
+  "modules_to_freeze": [
+    "vae_model_video",
+    "vae_model_audio",
+    "text_model",
+    "fusion_model",
+    "mllm.mllm_backbone",
+    "mllm.connector"
+  ],
+  "modules_to_unfreeze": [],
+  "negative_text_prompt": "",
+  "noise_scheduler_id": "Lightricks/LTX-Video",
+  "num_metaqueries": 128,
+  "ovi_gradient_checkpointing_block_count": null,
+  "prev_temp_chunk": 1,
+  "real_score_lora_checkpointing": {},
+  "rollout": {},
+  "scheduler_id": "Lightricks/LTX-Video",
+  "stage": 1,
+  "stage_name": null,
+  "student_checkpoint": null,
+  "student_lora_checkpointing": {},
+  "system_prompt": "You are a speech-video clip narrator for a text-to-audio-video (T2AV) model.\nThe input includes audio-visual dialogue between the Interlocutor and the Agent. The dialogue context is provided ONLY to infer what occurs next and must NOT be described, referenced, or summarized explicitly.\nYour task is to write a narration describing the NEXT recorded speech and video clip that shows ONLY the AGENT. Follow these rules precisely:\nVISUAL RULES: - Provide a HIGH-DETAIL appearance sentence, covering: (a) Gender + precise clothing (material/color/style, e.g., \"velvet head covering with ruffled accent\"), (b) Distinctive features (eyeliner/beard/accessories), (c) Scene composition (framing/lighting/color tone/backdrop, e.g., \"stark black and white lighting against plain curtains\"). - Describe ONLY visible physical actions in strict chronological order: (a) Facial muscles (e.g., \"lips curve upward forming a subtle smile\"), (b) Eye movements (e.g., \"gaze drifts downward for two seconds\"), (c) Posture/gestures (e.g., \"thumb brushes beard thoughtfully\"). - Emotions MUST derive from observable cues: CORRECT: \"eyes crinkling at the corners in a warm smile\" (visible) WRONG: \"she feels happy\" (internal) or \"looks thoughtful\" (abstract).\nSPEECH RULES: - Determine speech occurrence based on PROVIDED TARGETS or NATURAL CONVERSATION FLOW: - IF TARGETS ARE PROVIDED: (a) If \"[Target State]: Listening\" is provided in the context, describe silent actions ONLY without any speech content and do not use <S><E> tags. (b) If \"[Target Speech]: [Text]\" is provided, use the provided [Text] exactly as full speech content. - IF TARGETS ARE NOT PROVIDED: (a) If dialogue context implies continued listening or non-verbal reaction, describe silent actions ONLY without any speech content and do not use <S><E> tags. (b) If dialogue context implies the agent WILL SPEAK NEXT (even after brief listening actions), generate speech. (c) When speech needs to be generated: (1) Infer the speech content EXCLUSIVELY from dialogue context. (2) The speech content must be a RESPONSE that move forward the conversation, but word overlap with the end of the agent\u2019s speech context is STRICTLY FORBIDDEN. - GENERAL SPEECH STYLE (Only for speech existing in the narration): (b) Wrap each segment in <S> and <E> (speech contents are broken down into multiple segments when there are too many words for one segment) in the MIDDLE of the narration. (c) For each speech segment, embed it NATURALLY within action descriptions BEFORE and AFTER speech: CORRECT: \"meeting the camera, <S>I have considered your proposal<E> he brushes his thumb against his beard.\" WRONG: separate speech/action blocks.\nAUDIO RULES: - Add \"Audio: [description]\" at the VERY END of the paragraph, separated by a single space. - If no speech occurs (means no <S><E> tags exist), use \"SILENT\" for the audio description. - Otherwise, describe ONLY the agent\u2019s audio characteristics: (a) Voice traits: Describe the voice of the AGENT only. If the agent has spoken in the prior context, describe that voice in detail (pitch/tone/pace/texture, e.g., \"low-pitched with slight rasp\"). If the agent has NOT yet spoken in the context, you MUST ignore the interlocutor\u2019s voice entirely and infer a new, plausible voice for the AGENT EXCLUSIVELY based on his visual appearance (age, gender, physique). (b) Emotional progression (e.g., \"reflective shifting to joyful tone\"). (c) Recording environment ONLY if it affects vocal quality (e.g., \"dry and direct, quiet room\"). - NEVER mention background sounds, other voices, or specific locations (e.g., \"gift shop\"). - Never mix or attribute the interlocutor\u2019s vocal traits to the agent.\nTEMPORAL & STYLE RULES: - Chronological flow: [Full appearance sentence] \u2192 [Pre-speech actions] \u2192 [<S>Speech Content<E>] \u2192 [Post-speech actions] \u2192 (repeat for multiple speech segments if any) \u2192 Audio line - Appearance sentence must start with appearance-based identifier (e.g., \"A woman in a blue dress...\"), then use pronouns (she/he/they). - Use PRESENT TENSE throughout (e.g., \"she wears\", \"he gestures\"). - Write in THIRD-PERSON, single continuous paragraph. NO bullet points, headings, or line breaks. - NEVER use: agent, assistant, interlocutor, respond, reply, listen, understand.\nOUTPUT EXAMPLES (CRITICAL FOR FORMATTING): EXAMPLE 1 (Natural pause before speech, no target provided): \"A man in a grey sweatshirt stands before dark red curtains. His eyes track the interlocutor for two seconds, meeting the camera with a steady gaze, <S>I have been considering your proposal all week.<E> straightening his shoulders as his thumb brushes his beard thoughtfully. Audio: Calm low-pitched male voice with steady rhythm and slight rasp, recorded in a quiet room.\"\nEXAMPLE 2 (Silent reaction only, no target provided): \"A woman in a dark velvet head covering sits in stark black and white lighting. Her gaze drifts downward for three seconds, eyelids lowering slowly. Fingers tap once on the armrest as her head tilts upward with a subtle smile, eyes crinkling at the corners while maintaining silence. Audio: SILENT.\"\nEXAMPLE 3 (Multiple speech segments with emotional progression, no target provided):\nEXAMPLE 4 (Target provided and speech break into multiple segments, [Target Speech]: \"I completely understand your frustration with the current situation, but we need to stay focused on the long-term goals of the project.\"): \"A woman in a silk blouse sits against a bright window. She nods slowly as the interlocutor finishes, then takes a deep breath. Meeting the camera with a steady gaze, <S>I completely understand your frustration with the current situation,<E> her expression softens. Pausing briefly to adjust her posture, she continues, <S>but we need to stay focused on the long-term goals of the project.<E> She remains attentive, her eyes reflecting deep thought while awaiting a reply. Audio: Articulate female voice with a calm, reassuring tone and steady pace, recorded in a quiet environment.\"\nEXAMPLE 5 (Target provided andcontains [Target State]: Listening): \"A man in a flannel shirt leans back against a brick wall. He tilts his head slightly to the right, eyes tracking the interlocutor\u2019s movements. His brow furrows momentarily in concentration, lips pressed together in a neutral line. He maintains a steady, receptive posture, nodding once in acknowledgement while remaining silent. Audio: SILENT.\"\nNEVER OUTPUT LIKE THIS (common failures): - \"...last words of agent\u2019s context speech. <S>Last words of context, then new content.<E>\" // Repeating context - \"The agent looks thoughtful after hearing the question.\"  // Banned words + internal state - \"<S>I agree<E> <S>But I don\u2019t think it\u2019s a good idea<E>\"  // Short speech broken down into multiple segments and not embedded in actions - \"Audio: Quiet room with distant traffic.\"  // Illegal background sound - \"A man speaks about his recovery.\"  // Invented topic (must derive from context) - \"She has a kind face.\"  // Abstract trait without visible cues",
+  "target_same_timestep_across_blocks": false,
+  "teacher_checkpoint": null,
+  "teacher_lora_checkpointing": {},
+  "text_encoder_id": "./pretrained_models/Wan2.2-TI2V-5B",
+  "text_encoder_path": "./pretrained_models/Wan2.2-TI2V-5B/models_t5_umt5-xxl-enc-bf16.pth",
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.3",
+  "use_ref_as_sink": true,
+  "use_ref_audio_concat": {
+    "enable": true,
+    "uncond_ratio": 0.1
+  },
+  "vae_downsample_f": 16,
+  "vae_downsample_t": 4,
+  "vae_id": "Lightricks/LTX-Video",
+  "vae_path": "./pretrained_models/Wan2.2-TI2V-5B/Wan2.2_VAE.pth",
+  "video_context_noise": 0.5,
+  "video_lora_dict": {
+    "adapter_name": "s2_ovi_vid_lora",
+    "enable": true,
+    "lora_dropout": 0.0,
+    "network_alpha": 128,
+    "rank": 128,
+    "target_modules": [
+      "q",
+      "k",
+      "v",
+      "o",
+      "ffn.0",
+      "ffn.2"
+    ]
+  },
+  "video_negative_prompt": "jitter, bad hands, blur, distortion",
+  "window": {}
+}

backup-260717/0220_stage2_ar_tf_frame3_context_noise_0p5_aud_lookahead_a2a_all_v2a_1-checkpoint-8800/latest

@@ -0,0 +1 @@
+global_step8800

backup-260717/0220_stage2_ar_tf_frame3_context_noise_0p5_aud_lookahead_a2a_all_v2a_1-checkpoint-8800/zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if ZERO_STAGE not in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info("Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info("Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)

backup-260717/0222_stage3_causal_cd_frame3-checkpoint-5200/config.json

@@ -0,0 +1,246 @@
+{
+  "_gradient_checkpointing": true,
+  "ar_state_text_system_prompt": "You are a helpful assistant, capable of perceiving auditory and textual inputs, as well as generating text responses.",
+  "ar_target_debug": true,
+  "ar_target_do_sample": true,
+  "ar_target_max_new_tokens": 64,
+  "ar_target_temperature": 0.9,
+  "ar_target_top_k": 50,
+  "ar_target_top_p": 0.9,
+  "ar_target_use_v2": true,
+  "architectures": [
+    "Stage3CausalCDMetaQuery"
+  ],
+  "attn_implementation": "auto",
+  "audio_context_noise": 0.5,
+  "audio_lora_dict": {
+    "adapter_name": "s2_ovi_audio_lora",
+    "enable": true,
+    "lora_dropout": 0.0,
+    "network_alpha": 128,
+    "rank": 128,
+    "target_modules": [
+      "q",
+      "k",
+      "v",
+      "o",
+      "ffn.0",
+      "ffn.2"
+    ]
+  },
+  "audio_loss_weight": 0.5,
+  "audio_negative_prompt": "robotic, muffled, echo, distorted",
+  "audio_vae_path": "./pretrained_models/MMAudio/ext_weights/",
+  "block_causal": {
+    "a2a_lookahead": -1,
+    "attention_backend": "auto",
+    "audio_kv_cache_size": 4,
+    "enable": true,
+    "frame_per_block": 3,
+    "teacher_forcing": true,
+    "v2a_lookahead": 1,
+    "video_kv_cache_size": 4
+  },
+  "bsmntw_weighting": true,
+  "connector_num_hidden_layers": 8,
+  "diffusion_forcing": {
+    "enable": false,
+    "prefix_only": false,
+    "prob": 0.9,
+    "start_step": 0,
+    "temp_chunk": 3
+  },
+  "diffusion_model_id": "Lightricks/LTX-Video",
+  "diffusion_model_path": "./pretrained_models/Ovi/model_960x960_10s.safetensors",
+  "distillation": {
+    "audio_loss_weight": 0.5,
+    "audio_scale": 4.5,
+    "denoising_step_lists": {
+      "1": [
+        1000
+      ],
+      "2": [
+        1000,
+        500
+      ],
+      "3": [
+        1000,
+        750,
+        500
+      ],
+      "4": [
+        1000,
+        750,
+        500,
+        250
+      ]
+    },
+    "denoising_step_lists_first_chunk": {
+      "1": [
+        1000,
+        750,
+        500,
+        250
+      ],
+      "2": [
+        1000,
+        750,
+        500,
+        250
+      ],
+      "3": [
+        1000,
+        750,
+        500,
+        250
+      ],
+      "4": [
+        1000,
+        750,
+        500,
+        250
+      ]
+    },
+    "detach_mllm_mq": true,
+    "discrete_cd_n": 48,
+    "ema_decay": 0.99,
+    "flowmatch_shift": 5.0,
+    "guidance_scale": 3.5,
+    "id_scale": 4.5,
+    "listening_loss_weight": 1.0,
+    "num_inference_steps": 4,
+    "num_train_timestep": 1000,
+    "slg_layer": 11,
+    "trainer": "causal_consistency_distillation",
+    "video_loss_weight": 1.0,
+    "warp_denoising_step": true
+  },
+  "dmd": {},
+  "fake_score_learning_rate": null,
+  "fake_score_lora_checkpointing": {},
+  "flowmatch_shift": 5.0,
+  "gen_fps": 25.0,
+  "generator_lora_checkpointing": {},
+  "has_context_prompt": {
+    "enable": true,
+    "gt_t5_transcript_ratio": 1.0,
+    "no_mq_context_uncond_ratio": 0.0,
+    "skip_mllm_encode_ratio": 0.3
+  },
+  "in_channels": 16,
+  "infer": false,
+  "latent_spatial_size": 28,
+  "latent_temporal_size": 25,
+  "listening_loss_weight": 1.0,
+  "listening_prompt": "not speaking, closed mouth, listening, silent",
+  "local_files_only": true,
+  "lora_checkpointing": {
+    "audio_adapter_name": "s2_ovi_audio_lora",
+    "merge_from_checkpoints": [
+      {
+        "load_non_lora": true,
+        "path": "./output/0217_s1_qwen2p5omni7b_ovi1p1_mq_nonverbal_context_and_T5_response_talking80_vision_audio_loss_0p5_ICL_voice_prefix_sink_loss"
+      }
+    ],
+    "merge_load_non_lora": true,
+    "resume_from_checkpoint": "./output/0220_stage2_ar_tf_frame3_context_noise_0p5_aud_lookahead_a2a_all_v2a_1",
+    "resume_lora_only": true,
+    "video_adapter_name": "s2_ovi_vid_lora"
+  },
+  "lora_dict": {},
+  "max_input_text_tokens": 256,
+  "mllm_ce_only": false,
+  "mllm_id": "./pretrained_models/Qwen2.5-Omni-7B",
+  "mllm_local_path": null,
+  "mllm_lora_dict": {
+    "alpha": 64,
+    "dropout": 0.0,
+    "enable": false,
+    "rank": 32,
+    "target_modules": [
+      "q_proj",
+      "k_proj",
+      "v_proj",
+      "o_proj",
+      "gate_proj",
+      "up_proj",
+      "down_proj"
+    ]
+  },
+  "model_type": "stage3_causal_cd_metaquery",
+  "modules_to_freeze": [
+    "vae_model_video",
+    "vae_model_audio",
+    "text_model",
+    "fusion_model",
+    "mllm.mllm_backbone",
+    "mllm.connector"
+  ],
+  "modules_to_unfreeze": [],
+  "negative_text_prompt": "",
+  "noise_scheduler_id": "Lightricks/LTX-Video",
+  "num_metaqueries": 128,
+  "ovi_gradient_checkpointing_block_count": null,
+  "prev_temp_chunk": 1,
+  "real_score_lora_checkpointing": {},
+  "rollout": {},
+  "scheduler_id": "Lightricks/LTX-Video",
+  "stage": 3,
+  "stage_name": "causal_cd",
+  "student_checkpoint": null,
+  "student_lora_checkpointing": {
+    "audio_adapter_name": "s2_ovi_audio_lora",
+    "merge_from_checkpoints": [
+      {
+        "load_non_lora": true,
+        "path": "./output/0217_s1_qwen2p5omni7b_ovi1p1_mq_nonverbal_context_and_T5_response_talking80_vision_audio_loss_0p5_ICL_voice_prefix_sink_loss"
+      }
+    ],
+    "merge_load_non_lora": true,
+    "resume_from_checkpoint": "./output/0220_stage2_ar_tf_frame3_context_noise_0p5_aud_lookahead_a2a_all_v2a_1",
+    "resume_lora_only": true,
+    "video_adapter_name": "s2_ovi_vid_lora"
+  },
+  "system_prompt": "You are a speech-video clip narrator for a text-to-audio-video (T2AV) model.\nThe input includes audio-visual dialogue between the Interlocutor and the Agent. The dialogue context is provided ONLY to infer what occurs next and must NOT be described, referenced, or summarized explicitly.\nYour task is to write a narration describing the NEXT recorded speech and video clip that shows ONLY the AGENT. Follow these rules precisely:\nVISUAL RULES: - Provide a HIGH-DETAIL appearance sentence, covering: (a) Gender + precise clothing (material/color/style, e.g., \"velvet head covering with ruffled accent\"), (b) Distinctive features (eyeliner/beard/accessories), (c) Scene composition (framing/lighting/color tone/backdrop, e.g., \"stark black and white lighting against plain curtains\"). - Describe ONLY visible physical actions in strict chronological order: (a) Facial muscles (e.g., \"lips curve upward forming a subtle smile\"), (b) Eye movements (e.g., \"gaze drifts downward for two seconds\"), (c) Posture/gestures (e.g., \"thumb brushes beard thoughtfully\"). - Emotions MUST derive from observable cues: CORRECT: \"eyes crinkling at the corners in a warm smile\" (visible) WRONG: \"she feels happy\" (internal) or \"looks thoughtful\" (abstract).\nSPEECH RULES: - Determine speech occurrence based on PROVIDED TARGETS or NATURAL CONVERSATION FLOW: - IF TARGETS ARE PROVIDED: (a) If \"[Target State]: Listening\" is provided in the context, describe silent actions ONLY without any speech content and do not use <S><E> tags. (b) If \"[Target Speech]: [Text]\" is provided, use the provided [Text] exactly as full speech content. - IF TARGETS ARE NOT PROVIDED: (a) If dialogue context implies continued listening or non-verbal reaction, describe silent actions ONLY without any speech content and do not use <S><E> tags. (b) If dialogue context implies the agent WILL SPEAK NEXT (even after brief listening actions), generate speech. (c) When speech needs to be generated: (1) Infer the speech content EXCLUSIVELY from dialogue context. (2) The speech content must be a RESPONSE that move forward the conversation, but word overlap with the end of the agent\u2019s speech context is STRICTLY FORBIDDEN. - GENERAL SPEECH STYLE (Only for speech existing in the narration): (b) Wrap each segment in <S> and <E> (speech contents are broken down into multiple segments when there are too many words for one segment) in the MIDDLE of the narration. (c) For each speech segment, embed it NATURALLY within action descriptions BEFORE and AFTER speech: CORRECT: \"meeting the camera, <S>I have considered your proposal<E> he brushes his thumb against his beard.\" WRONG: separate speech/action blocks.\nAUDIO RULES: - Add \"Audio: [description]\" at the VERY END of the paragraph, separated by a single space. - If no speech occurs (means no <S><E> tags exist), use \"SILENT\" for the audio description. - Otherwise, describe ONLY the agent\u2019s audio characteristics: (a) Voice traits: Describe the voice of the AGENT only. If the agent has spoken in the prior context, describe that voice in detail (pitch/tone/pace/texture, e.g., \"low-pitched with slight rasp\"). If the agent has NOT yet spoken in the context, you MUST ignore the interlocutor\u2019s voice entirely and infer a new, plausible voice for the AGENT EXCLUSIVELY based on his visual appearance (age, gender, physique). (b) Emotional progression (e.g., \"reflective shifting to joyful tone\"). (c) Recording environment ONLY if it affects vocal quality (e.g., \"dry and direct, quiet room\"). - NEVER mention background sounds, other voices, or specific locations (e.g., \"gift shop\"). - Never mix or attribute the interlocutor\u2019s vocal traits to the agent.\nTEMPORAL & STYLE RULES: - Chronological flow: [Full appearance sentence] \u2192 [Pre-speech actions] \u2192 [<S>Speech Content<E>] \u2192 [Post-speech actions] \u2192 (repeat for multiple speech segments if any) \u2192 Audio line - Appearance sentence must start with appearance-based identifier (e.g., \"A woman in a blue dress...\"), then use pronouns (she/he/they). - Use PRESENT TENSE throughout (e.g., \"she wears\", \"he gestures\"). - Write in THIRD-PERSON, single continuous paragraph. NO bullet points, headings, or line breaks. - NEVER use: agent, assistant, interlocutor, respond, reply, listen, understand.\nOUTPUT EXAMPLES (CRITICAL FOR FORMATTING): EXAMPLE 1 (Natural pause before speech, no target provided): \"A man in a grey sweatshirt stands before dark red curtains. His eyes track the interlocutor for two seconds, meeting the camera with a steady gaze, <S>I have been considering your proposal all week.<E> straightening his shoulders as his thumb brushes his beard thoughtfully. Audio: Calm low-pitched male voice with steady rhythm and slight rasp, recorded in a quiet room.\"\nEXAMPLE 2 (Silent reaction only, no target provided): \"A woman in a dark velvet head covering sits in stark black and white lighting. Her gaze drifts downward for three seconds, eyelids lowering slowly. Fingers tap once on the armrest as her head tilts upward with a subtle smile, eyes crinkling at the corners while maintaining silence. Audio: SILENT.\"\nEXAMPLE 3 (Multiple speech segments with emotional progression, no target provided):\nEXAMPLE 4 (Target provided and speech break into multiple segments, [Target Speech]: \"I completely understand your frustration with the current situation, but we need to stay focused on the long-term goals of the project.\"): \"A woman in a silk blouse sits against a bright window. She nods slowly as the interlocutor finishes, then takes a deep breath. Meeting the camera with a steady gaze, <S>I completely understand your frustration with the current situation,<E> her expression softens. Pausing briefly to adjust her posture, she continues, <S>but we need to stay focused on the long-term goals of the project.<E> She remains attentive, her eyes reflecting deep thought while awaiting a reply. Audio: Articulate female voice with a calm, reassuring tone and steady pace, recorded in a quiet environment.\"\nEXAMPLE 5 (Target provided andcontains [Target State]: Listening): \"A man in a flannel shirt leans back against a brick wall. He tilts his head slightly to the right, eyes tracking the interlocutor\u2019s movements. His brow furrows momentarily in concentration, lips pressed together in a neutral line. He maintains a steady, receptive posture, nodding once in acknowledgement while remaining silent. Audio: SILENT.\"\nNEVER OUTPUT LIKE THIS (common failures): - \"...last words of agent\u2019s context speech. <S>Last words of context, then new content.<E>\" // Repeating context - \"The agent looks thoughtful after hearing the question.\"  // Banned words + internal state - \"<S>I agree<E> <S>But I don\u2019t think it\u2019s a good idea<E>\"  // Short speech broken down into multiple segments and not embedded in actions - \"Audio: Quiet room with distant traffic.\"  // Illegal background sound - \"A man speaks about his recovery.\"  // Invented topic (must derive from context) - \"She has a kind face.\"  // Abstract trait without visible cues",
+  "target_same_timestep_across_blocks": false,
+  "teacher_checkpoint": null,
+  "teacher_lora_checkpointing": {
+    "audio_adapter_name": "s2_ovi_audio_lora",
+    "merge_from_checkpoint": "./output/0217_s1_qwen2p5omni7b_ovi1p1_mq_nonverbal_context_and_T5_response_talking80_vision_audio_loss_0p5_ICL_voice_prefix_sink_loss",
+    "merge_load_non_lora": true,
+    "resume_from_checkpoint": "./output/0220_stage2_ar_tf_frame3_context_noise_0p5_aud_lookahead_a2a_all_v2a_1",
+    "resume_lora_only": true,
+    "video_adapter_name": "s2_ovi_vid_lora"
+  },
+  "text_encoder_id": "./pretrained_models/Wan2.2-TI2V-5B",
+  "text_encoder_path": "./pretrained_models/Wan2.2-TI2V-5B/models_t5_umt5-xxl-enc-bf16.pth",
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.3",
+  "use_ref_as_sink": true,
+  "use_ref_audio_concat": {
+    "enable": true,
+    "uncond_ratio": 0.0
+  },
+  "vae_downsample_f": 16,
+  "vae_downsample_t": 4,
+  "vae_id": "Lightricks/LTX-Video",
+  "vae_path": "./pretrained_models/Wan2.2-TI2V-5B/Wan2.2_VAE.pth",
+  "video_context_noise": 0.5,
+  "video_lora_dict": {
+    "adapter_name": "s2_ovi_vid_lora",
+    "enable": true,
+    "lora_dropout": 0.0,
+    "network_alpha": 128,
+    "rank": 128,
+    "target_modules": [
+      "q",
+      "k",
+      "v",
+      "o",
+      "ffn.0",
+      "ffn.2"
+    ]
+  },
+  "video_negative_prompt": "jitter, bad hands, blur, distortion",
+  "window": {}
+}

backup-260717/0222_stage3_causal_cd_frame3-checkpoint-5200/latest

@@ -0,0 +1 @@
+global_step5200

backup-260717/0222_stage3_causal_cd_frame3-checkpoint-5200/zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if ZERO_STAGE not in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info("Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info("Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)