Upload checkpoints/step-00500 (k2_apex_ctx)

checkpoints/step-00500/adapter_config.json

@@ -0,0 +1,52 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "monteclora_config": null,
+  "peft_type": "LORA",
+  "peft_version": "0.20.0",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "up",
+    "gate",
+    "txtmlp.1",
+    "wk",
+    "txtmlp.3",
+    "down",
+    "wv",
+    "wq",
+    "wo"
+  ],
+  "target_parameters": null,
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false,
+  "velora_config": null
+}

checkpoints/step-00500/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0ce2ffccd48e4c4bb76c39f560d1c89e39fa60ee7ec581c3e9f44ee0dfc4639
+size 459800344

checkpoints/step-00500/apex_ctx_512.toml

@@ -0,0 +1,66 @@
+# krea2_apex @512 — CTX (docs/KREA2_APEX_SPEC.md §2.7).
+#
+# Contrato (o que difere de todo braço anterior):
+#   - geometria frame_fit: ref no eixo de frame (pos[...,0]=1) e CENTRADA no
+#     grid do alvo com offsets FRACIONÁRIOS; sem width_shift, sem reescala
+#   - ref preprocessada com o ramo `fit` do comfyui-krea2edit (CROP_TOL 0.08,
+#     /16 floor com cap, crop-to-grid, bicubic+antialias)
+#   - reference_timestep='target' (o node calcula UM tvec para toda a sequência;
+#     um LoRA t=0 degrada lá em silêncio)
+#   - timestep_law='krea2_apex': t = sigmoid(s(N)·z + m(N)), peso da loss ≡ 1,
+#     amostragem estratificada golden-ratio. MUTUAMENTE EXCLUSIVO com flux_shift
+#   - LoRA em blocks + txtfusion + txtmlp (258 módulos)
+#   - live_text_encoding: sem cache de texto; jitter de grounding e clause
+#     dropout re-sorteados POR STEP
+
+output_dir = '/workspace/checkpoints/k2_apex_ctx'
+dataset = '/workspace/configs/apex_ctx_dataset.toml'
+epochs = 1000
+max_steps = 5000
+micro_batch_size_per_gpu = 1
+gradient_accumulation_steps = 1
+pipeline_stages = 1
+gradient_clipping = 1.0
+warmup_steps = 250
+save_every_n_steps = 250
+checkpoint_every_n_minutes = 120
+activation_checkpointing = true
+save_dtype = 'bfloat16'
+caching_batch_size = 1       # OBRIGATÓRIO: o fit dá refs de shapes diferentes
+blocks_to_swap = 0
+
+[model]
+type = 'krea2_apex'
+diffusion_model = '/workspace/models/krea2/diffusion_models/krea2_raw_fp8_scaled.safetensors'
+vae = '/workspace/models/krea2/split_files/vae/qwen_image_vae.safetensors'
+text_encoders = [
+    {path = '/workspace/models/krea2/text_encoders/qwen3vl_4b_bf16.safetensors', type = 'krea2'},
+]
+dtype = 'bfloat16'
+diffusion_model_dtype = 'float8'
+timestep_sample_method = 'logit_normal'
+timestep_law = 'krea2_apex'   # NÃO usar flux_shift/shift junto
+live_text_encoding = true
+
+[krea2_apex]
+reference_timestep = 'target'
+vl_prompt_style = 'plain'
+vl_longest_side = 768
+vl_grounding_jitter = [384, 768]   # agora POR STEP
+caption_dropout = 0.05
+clause_dropout_prob = 0.40
+clause_dropout_rate = 0.60
+keep_clauses = 1
+
+[adapter]
+type = 'lora'
+rank = 64
+# alpha: o trainer força alpha = rank (scaling 1). Não declarar.
+dtype = 'bfloat16'
+
+[optimizer]
+type = 'AdamW8bitKahan'
+lr = 1e-4
+betas = [0.9, 0.99]
+weight_decay = 0.01
+stabilize = false