diff --git a/src/maxdiffusion/configs/base_flux2klein.yml b/src/maxdiffusion/configs/base_flux2klein.yml index f2813c8fd..277712a2f 100644 --- a/src/maxdiffusion/configs/base_flux2klein.yml +++ b/src/maxdiffusion/configs/base_flux2klein.yml @@ -75,6 +75,17 @@ mask_padding_tokens: True attention_sharding_uniform: True flash_block_sizes: {} +block_q: 0 +block_kv: 0 +block_kv_compute: 0 +ulysses_shards: -1 +ulysses_attention_chunks: 1 +text_encoder_attention: 'dot_product' +text_encoder_flash_block_sizes: {} +text_encoder_block_q: 0 +text_encoder_block_kv: 0 +text_encoder_block_kv_compute: 0 +text_encoder_max_layer: 27 # GroupNorm groups norm_num_groups: 32 @@ -154,12 +165,12 @@ data_sharding: [['data', 'fsdp', 'context', 'tensor']] # value to auto-shard based on available slices and devices. # By default, product of the DCN axes should equal number of slices # and product of the ICI axes should equal number of devices per slice. -dcn_data_parallelism: 1 # recommended DCN axis to be auto-sharded -dcn_fsdp_parallelism: -1 +dcn_data_parallelism: 1 +dcn_fsdp_parallelism: 1 dcn_context_parallelism: 1 dcn_tensor_parallelism: 1 ici_data_parallelism: 1 -ici_fsdp_parallelism: -1 +ici_fsdp_parallelism: 1 ici_context_parallelism: 1 ici_tensor_parallelism: 1 @@ -203,7 +214,7 @@ num_train_epochs: 1 seed: 0 output_dir: 'output/' output_name: "flux2klein_generated_image.png" -per_device_batch_size: 1 +per_device_batch_size: 1.0 warmup_steps_fraction: 0.1 learning_rate_schedule_steps: -1 # By default the length of the schedule is set to the number of steps. @@ -231,6 +242,7 @@ do_classifier_free_guidance: True guidance_scale: 4.0 guidance_rescale: 0.0 num_inference_steps: 4 +num_reps: 1 save_final_checkpoint: False # SDXL Lightning parameters diff --git a/src/maxdiffusion/configs/base_flux2klein_9B.yml b/src/maxdiffusion/configs/base_flux2klein_9B.yml index a6c670a69..1b1b55ec4 100644 --- a/src/maxdiffusion/configs/base_flux2klein_9B.yml +++ b/src/maxdiffusion/configs/base_flux2klein_9B.yml @@ -75,6 +75,17 @@ mask_padding_tokens: True attention_sharding_uniform: True flash_block_sizes: {} +block_q: 0 +block_kv: 0 +block_kv_compute: 0 +ulysses_shards: -1 +ulysses_attention_chunks: 1 +text_encoder_attention: 'dot_product' +text_encoder_flash_block_sizes: {} +text_encoder_block_q: 0 +text_encoder_block_kv: 0 +text_encoder_block_kv_compute: 0 +text_encoder_max_layer: 27 # GroupNorm groups norm_num_groups: 32 @@ -154,12 +165,12 @@ data_sharding: [['data', 'fsdp', 'context', 'tensor']] # value to auto-shard based on available slices and devices. # By default, product of the DCN axes should equal number of slices # and product of the ICI axes should equal number of devices per slice. -dcn_data_parallelism: 1 # recommended DCN axis to be auto-sharded -dcn_fsdp_parallelism: -1 +dcn_data_parallelism: 1 +dcn_fsdp_parallelism: 1 dcn_context_parallelism: 1 dcn_tensor_parallelism: 1 ici_data_parallelism: 1 -ici_fsdp_parallelism: -1 # recommended ICI axis to be auto-sharded +ici_fsdp_parallelism: 1 ici_context_parallelism: 1 ici_tensor_parallelism: 1 @@ -203,7 +214,7 @@ num_train_epochs: 1 seed: 0 output_dir: 'output/' output_name: "flux2klein_generated_image.png" -per_device_batch_size: 1 +per_device_batch_size: 1.0 warmup_steps_fraction: 0.1 learning_rate_schedule_steps: -1 # By default the length of the schedule is set to the number of steps. @@ -231,6 +242,7 @@ do_classifier_free_guidance: True guidance_scale: 4.0 guidance_rescale: 0.0 num_inference_steps: 4 +num_reps: 1 save_final_checkpoint: False # SDXL Lightning parameters diff --git a/src/maxdiffusion/generate_flux2klein.py b/src/maxdiffusion/generate_flux2klein.py index 7956c850d..996b782d4 100644 --- a/src/maxdiffusion/generate_flux2klein.py +++ b/src/maxdiffusion/generate_flux2klein.py @@ -40,6 +40,7 @@ from maxdiffusion.models.qwen3_flax import FlaxQwen3Config, FlaxQwen3Model from maxdiffusion.models.qwen3_utils import load_and_convert_qwen3_weights from maxdiffusion.schedulers.scheduling_flow_match_flax import FlaxFlowMatchScheduler +from maxdiffusion.pipelines.flux.flux2klein_pipeline import FlaxFlux2KleinPipeline def partition_prompts(prompt_str: str, batch_size: int) -> List[str]: @@ -79,8 +80,22 @@ def encode_prompt(prompt: str, snapshot_dir: str = None, repo_id: str = "black-f text_encoder_path = os.path.join(snapshot_dir, "text_encoder") tokenizer_path = os.path.join(snapshot_dir, "tokenizer") - if not os.path.exists(tokenizer_path): - tokenizer_path = text_encoder_path + + if not os.path.exists(os.path.join(text_encoder_path, "config.json")) or not os.path.exists(tokenizer_path): + try: + fb_dir = snapshot_download(repo_id=repo_id, local_files_only=True) + if not os.path.exists(os.path.join(text_encoder_path, "config.json")): + text_encoder_path = os.path.join(fb_dir, "text_encoder") + if not os.path.exists(tokenizer_path): + tokenizer_path = ( + os.path.join(fb_dir, "tokenizer") + if os.path.exists(os.path.join(fb_dir, "tokenizer")) + else os.path.join(fb_dir, "text_encoder") + ) + except Exception: + if not os.path.exists(tokenizer_path): + tokenizer_path = text_encoder_path + tokenizer = AutoTokenizer.from_pretrained(tokenizer_path) text_encoder = AutoModelForCausalLM.from_pretrained(text_encoder_path, torch_dtype=torch.float32) text_encoder.eval() @@ -134,20 +149,45 @@ def main(argv): from maxdiffusion.models.flux.util import ( load_and_convert_flux_klein_weights, load_and_convert_vae_weights, - cast_dict_to_bfloat16_inplace, ) - from maxdiffusion.pipelines.flux.flux2klein_pipeline import FlaxFlux2KleinPipeline config = pyconfig.config os.makedirs(config.output_dir, exist_ok=True) + if hasattr(config, "per_device_batch_size") and config.per_device_batch_size > 0: + calculated_batch_size = int(config.per_device_batch_size * jax.device_count()) + assert calculated_batch_size >= 1, ( + f"Calculated global batch_size is {calculated_batch_size}, which is invalid (must be >= 1). " + f"per_device_batch_size={config.per_device_batch_size} multiplied by jax.device_count()={jax.device_count()} " + f"evaluated to {config.per_device_batch_size * jax.device_count()}, which truncates to 0. " + f"Please increase per_device_batch_size or specify an explicit batch_size in your configuration." + ) + if calculated_batch_size != config.batch_size: + max_logging.log( + f"ℹ️ Updating batch_size from {config.batch_size} to {calculated_batch_size} " + f"based on per_device_batch_size={config.per_device_batch_size} and device_count={jax.device_count()}." + ) + pyconfig._config.keys["batch_size"] = calculated_batch_size + # 2. Setup device mesh - if config.batch_size == 1 and config.ici_tensor_parallelism == 1 and jax.device_count() > 1: + custom_parallelism_set = any( + any(arg.startswith(f"{k}=") for arg in sys.argv) + for k in [ + "ici_data_parallelism", + "ici_fsdp_parallelism", + "ici_context_parallelism", + "ici_tensor_parallelism", + ] + ) + + if not custom_parallelism_set and jax.device_count() > 1: max_logging.log( - f"ℹ️ Auto-configuring Tensor Parallelism: ici_tensor_parallelism={jax.device_count()}, ici_fsdp_parallelism=1 for batch_size=1 on {jax.device_count()} TPU devices." + f"ℹ️ Defaulting to Tensor Parallelism: ici_tensor_parallelism={jax.device_count()} on {jax.device_count()} TPU devices." ) pyconfig._config.keys["ici_tensor_parallelism"] = jax.device_count() + pyconfig._config.keys["ici_data_parallelism"] = 1 pyconfig._config.keys["ici_fsdp_parallelism"] = 1 + pyconfig._config.keys["ici_context_parallelism"] = 1 max_logging.log("Setting up JAX device mesh...") devices_array = create_device_mesh(config) @@ -174,8 +214,7 @@ def main(argv): # 3. Resolve weights repository snapshots repo_id = getattr(config, "pretrained_model_name_or_path", None) if not repo_id: - depth_val = getattr(config, "depth", None) - repo_id = "black-forest-labs/FLUX.2-klein-9B" if depth_val == 24 else "black-forest-labs/FLUX.2-klein-4B" + raise ValueError("pretrained_model_name_or_path must be specified in configuration YAML or CLI.") max_logging.log(f"Target model detected: {repo_id}") if os.path.exists(repo_id): @@ -184,8 +223,13 @@ def main(argv): else: from huggingface_hub import snapshot_download - max_logging.log(f"Resolving snapshot directory for model '{repo_id}' from HF Hub...") - snapshot_dir = snapshot_download(repo_id=repo_id) + rev = getattr(config, "revision", None) + if not rev or rev == "refs/pr/95": + rev = "main" + try: + snapshot_dir = snapshot_download(repo_id=repo_id, revision=rev, local_files_only=True) + except Exception: + snapshot_dir = snapshot_download(repo_id=repo_id, revision=rev) max_logging.log(f"Host {jax.process_index()} using HF snapshot directory: {snapshot_dir}") safetensors_path = os.path.join(snapshot_dir, "transformer") @@ -194,9 +238,33 @@ def main(argv): # 4. Load Qwen3 Config & Setup model layout from transformers import AutoConfig - - max_logging.log(f"Loading Qwen3 config from text_encoder path: {text_encoder_path}...") - pt_config = AutoConfig.from_pretrained(text_encoder_path, local_files_only=True) + from maxdiffusion.max_utils import get_flash_block_sizes + + pt_config = AutoConfig.from_pretrained(text_encoder_path) + + te_flash_block_sizes_dict = dict(getattr(config, "text_encoder_flash_block_sizes", {}) or {}) + for k in ["block_q", "block_kv", "block_kv_compute", "block_kv_compute_in", "heads_per_tile", "vmem_limit_bytes"]: + val = getattr(config, f"text_encoder_{k}", None) + if val is not None and val > 0: + te_flash_block_sizes_dict[k] = int(val) + if ( + "tokamax" in getattr(config, "text_encoder_attention", "") + or getattr(config, "text_encoder_attention", "") == "ulysses_ring" + ): + te_flash_block_sizes_dict.setdefault("block_q", 512) + te_flash_block_sizes_dict.setdefault("block_kv", 512) + te_flash_block_sizes_dict.setdefault("block_kv_compute", 512) + te_flash_block_sizes_dict.setdefault("block_q_dkv", te_flash_block_sizes_dict["block_q"]) + te_flash_block_sizes_dict.setdefault("block_kv_dkv", te_flash_block_sizes_dict["block_kv"]) + te_flash_block_sizes_dict.setdefault("block_kv_dkv_compute", te_flash_block_sizes_dict["block_kv_compute"]) + config.get_keys()["text_encoder_flash_block_sizes"] = te_flash_block_sizes_dict + + # Create a dummy config holder for max_utils.get_flash_block_sizes for Qwen3 + class _TEConfigHolder: + + def __init__(self, cfg): + self.flash_block_sizes = getattr(cfg, "text_encoder_flash_block_sizes", {}) + self.attention = getattr(cfg, "text_encoder_attention", "dot_product") qwen3_config = FlaxQwen3Config( vocab_size=pt_config.vocab_size, @@ -209,6 +277,12 @@ def main(argv): rms_norm_eps=pt_config.rms_norm_eps, rope_theta=pt_config.rope_theta, dtype=jnp.bfloat16 if config.weights_dtype == "bfloat16" else jnp.float32, + attention_kernel=getattr(config, "text_encoder_attention", "dot_product"), + flash_block_sizes=get_flash_block_sizes(_TEConfigHolder(config)), + mesh=mesh, + ulysses_shards=getattr(config, "ulysses_shards", -1), + ulysses_attention_chunks=getattr(config, "ulysses_attention_chunks", 1), + max_layer_to_run=getattr(config, "text_encoder_max_layer", 27), ) qwen3_model = FlaxQwen3Model(qwen3_config) @@ -217,9 +291,24 @@ def main(argv): transformer_config_json = os.path.join(safetensors_path, "config.json") transformer_pt_cfg = {} + loaded_cfg = False if os.path.exists(transformer_config_json): - with open(transformer_config_json, "r") as f: - transformer_pt_cfg = json.load(f) + try: + with open(transformer_config_json, "r") as f: + transformer_pt_cfg = json.load(f) + loaded_cfg = True + except Exception as e: + max_logging.log(f"ℹ️ Could not parse {transformer_config_json}: {e}. Falling back to HF cache...") + + if not loaded_cfg and repo_id: + try: + from huggingface_hub import hf_hub_download + + cfg_file = hf_hub_download(repo_id=repo_id, filename="transformer/config.json", local_files_only=True) + with open(cfg_file, "r") as f: + transformer_pt_cfg = json.load(f) + except Exception as e: + max_logging.log(f"⚠️ Warning resolving transformer config fallback from HF cache: {e}") num_double_layers = getattr(config, "num_double_layers", -1) if num_double_layers is None or num_double_layers <= 0: @@ -234,6 +323,27 @@ def main(argv): num_attention_heads = transformer_pt_cfg.get("num_attention_heads", 24) # 5. Instantiate JAX Flux2KleinTransformer2DModel + from maxdiffusion.max_utils import get_flash_block_sizes + + flash_block_sizes_val = getattr(config, "flash_block_sizes", {}) or {} + if isinstance(flash_block_sizes_val, str) and flash_block_sizes_val: + import ast + import json + + try: + flash_block_sizes_val = json.loads(flash_block_sizes_val) + except Exception: + try: + flash_block_sizes_val = ast.literal_eval(flash_block_sizes_val) + except Exception: + flash_block_sizes_val = {} + flash_block_sizes_dict = dict(flash_block_sizes_val) + for k in ["block_q", "block_kv", "block_kv_compute", "block_kv_compute_in", "heads_per_tile", "vmem_limit_bytes"]: + val = getattr(config, k, None) + if val is not None and val > 0: + flash_block_sizes_dict[k] = int(val) + config.get_keys()["flash_block_sizes"] = flash_block_sizes_dict + transformer = Flux2KleinTransformer2DModel( in_channels=128, num_layers=num_double_layers, @@ -255,6 +365,9 @@ def main(argv): dtype=jnp.bfloat16 if config.weights_dtype == "bfloat16" else jnp.float32, weights_dtype=jnp.bfloat16 if config.weights_dtype == "bfloat16" else jnp.float32, attention_kernel=config.attention, + flash_block_sizes=get_flash_block_sizes(config), + ulysses_shards=getattr(config, "ulysses_shards", -1), + ulysses_attention_chunks=getattr(config, "ulysses_attention_chunks", 1), scale_shift_order=getattr(config, "scale_shift_order", "shift_scale"), ) @@ -342,6 +455,7 @@ def qwen3_init_fn(): def unbox_fn(x): return x.unbox() if isinstance(x, flax_spmd.LogicallyPartitioned) else x + t_sub0 = time.time() params = jax.tree_util.tree_map( unbox_fn, abstract_transformer_vars["params"], is_leaf=lambda k: isinstance(k, flax_spmd.LogicallyPartitioned) ) @@ -357,17 +471,19 @@ def unbox_fn(x): ) qwen3_params = flax.core.unfreeze(qwen3_params) - params = load_and_convert_flux_klein_weights(safetensors_path, params, num_double_layers, depth) - vae_params, vae_bn_mean, vae_bn_std = load_and_convert_vae_weights(vae_safetensors_path, vae_params) - qwen3_params = load_and_convert_qwen3_weights(text_encoder_path, qwen3_params, qwen3_config) + max_logging.log(f" -> [SUB-TIMING 1/3] PyTree unboxing template setup: {time.time() - t_sub0:.2f}s") + t_sub1 = time.time() + + weight_dtype = jnp.bfloat16 if config.weights_dtype == "bfloat16" else jnp.float32 - if config.weights_dtype == "bfloat16": - max_logging.log("Casting JAX parameters to bfloat16 in-place...") - cast_dict_to_bfloat16_inplace(params, exclude_keywords=("norm",)) - cast_dict_to_bfloat16_inplace(vae_params, exclude_keywords=("norm",)) - cast_dict_to_bfloat16_inplace(qwen3_params, exclude_keywords=("norm",)) - vae_bn_mean = vae_bn_mean.astype(jnp.bfloat16) - vae_bn_std = vae_bn_std.astype(jnp.bfloat16) + params = load_and_convert_flux_klein_weights(safetensors_path, params, num_double_layers, depth, dtype=weight_dtype) + vae_params, vae_bn_mean, vae_bn_std = load_and_convert_vae_weights( + vae_safetensors_path, vae_params, dtype=weight_dtype + ) + qwen3_params = load_and_convert_qwen3_weights(text_encoder_path, qwen3_params, qwen3_config) + max_logging.log( + f" -> [SUB-TIMING 2/3] Safetensors loading & key mapping (in target dtype): {time.time() - t_sub1:.4f}s" + ) params = flax.core.freeze(params) vae_params = flax.core.freeze(vae_params) @@ -376,6 +492,7 @@ def unbox_fn(x): max_logging.log("\n" + "=" * 80) max_logging.log("🚀 Pinning all parameters to TPU HBM permanently...") max_logging.log("=" * 80 + "\n") + t_sub3 = time.time() max_logging.log("Putting params on TPU HBM...") with mesh, nn_partitioning.axis_rules(config.logical_axis_rules): try: @@ -394,12 +511,13 @@ def unbox_fn(x): vae_params = jax.tree_util.tree_map(max_utils.device_put_replicated, vae_params, vae_shardings) max_logging.log("Putting qwen3_params on TPU HBM...") qwen3_params = jax.tree_util.tree_map(max_utils.device_put_replicated, qwen3_params, qwen3_shardings) + max_logging.log(f" -> [SUB-TIMING 3/3] TPU HBM device_put placement: {time.time() - t_sub3:.4f}s") max_logging.log("All parameters placed on TPU HBM successfully!") gc.collect() jax.effects_barrier() load_time = time.time() - t_load_start - max_logging.log(f" -> [TIMING] Total Model Loading & Device Placement: {load_time:.2f} seconds ⏱️\n") + max_logging.log(f" -> [TIMING] Total Model Loading & Device Placement: {load_time:.4f} seconds ⏱️\n") # 9. Setup FlowMatch Scheduler scheduler = FlaxFlowMatchScheduler( @@ -426,16 +544,19 @@ def unbox_fn(x): mesh=mesh, ) - active_prompts = partition_prompts(config.prompt, config.batch_size) + prompt_str = getattr(config, "prompt", None) + if not prompt_str: + raise ValueError("Prompt must be specified in the configuration YAML or passed via CLI prompt='...'") + active_prompts = partition_prompts(prompt_str, config.batch_size) if getattr(config, "interactive", False): - print("\n" + "=" * 80) - print(" BATCHED INTERACTIVE GENERATION MODE ENABLED 🎮") - print("The model has been fully loaded and compiled on the TPU.") - print(f"Batch size: {config.batch_size} parallel images.") - print("Enter prompts separated by '||' (e.g. A cute cat || A red car)") - print("Type 'exit' to quit.") - print("=" * 80) + max_logging.log("\n" + "=" * 80) + max_logging.log(" BATCHED INTERACTIVE GENERATION MODE ENABLED 🎮") + max_logging.log("The model has been fully loaded and compiled on the TPU.") + max_logging.log(f"Batch size: {config.batch_size} parallel images.") + max_logging.log("Enter prompts separated by '||' (e.g. A cute cat || A red car)") + max_logging.log("Type 'exit' to quit.") + max_logging.log("=" * 80) image_idx = 1 while True: @@ -481,37 +602,23 @@ def unbox_fn(x): max_logging.log(f" -> Custom latents shape: {latents_to_use.shape} | sum: {latents_to_use.sum():.6f}") max_logging.log("\n" + "=" * 80) - max_logging.log("🚀 Running initial dry run (Warmup Pass) to compile XLA graphs...") + max_logging.log("🚀 Pre-compiling XLA graphs concurrently (AOT Compilation)...") max_logging.log("=" * 80) - _, warmup_trace = pipeline( - prompt=active_prompts, + aot_time = pipeline.compile_aot_async( params=params, vae_params=vae_params, qwen3_params=qwen3_params, vae_bn_mean=vae_bn_mean, vae_bn_std=vae_bn_std, - transformer_shardings=transformer_shardings, - vae_shardings=vae_shardings, - qwen3_shardings=qwen3_shardings, + batch_size=config.batch_size, height=config.height, width=config.width, - num_inference_steps=config.num_inference_steps, - batch_size=config.batch_size, - use_latents=use_latents_flag, - latents=latents_to_use, - output_dir=config.output_dir, - output_name="flux2klein_warmup.png", - ) - warmup_time = ( - warmup_trace.get("prompt_encoding", 0.0) - + warmup_trace.get("denoise_loop", 0.0) - + warmup_trace.get("vae_decode", 0.0) ) max_logging.log("\n" + "=" * 80) - max_logging.log("⏱️ Running timed pass at full TPU speed...") + max_logging.log("🚀 Running initial dry run (Warmup Pass) to verify compiled graph execution...") max_logging.log("=" * 80) - _, main_trace = pipeline( + _, warmup_trace = pipeline( prompt=active_prompts, params=params, vae_params=vae_params, @@ -528,24 +635,113 @@ def unbox_fn(x): use_latents=use_latents_flag, latents=latents_to_use, output_dir=config.output_dir, - output_name=config.output_name, + output_name="flux2klein_warmup.png", + warmup=True, ) - main_time = ( - main_trace.get("prompt_encoding", 0.0) + main_trace.get("denoise_loop", 0.0) + main_trace.get("vae_decode", 0.0) + warmup_time = ( + warmup_trace.get("prompt_encoding", 0.0) + + warmup_trace.get("denoise_loop", 0.0) + + warmup_trace.get("vae_decode", 0.0) ) + num_reps = int(getattr(config, "num_reps", 1)) + max_logging.log("\n" + "=" * 80) + max_logging.log(f"⏱️ Running timed pass at full TPU speed (num_reps={num_reps})...") + max_logging.log("=" * 80) + + main_traces = [] + main_times = [] + + for rep in range(num_reps): + rep_str = f" [Rep {rep+1}/{num_reps}]" if num_reps > 1 else "" + if rep > 0: + max_logging.log(f"⏱️ Running timed pass{rep_str}...") + + if max_utils.profiler_enabled(config) and rep == 0: + max_logging.log(f"🚀 XProf / JAX Profiler active! Capturing trace into: {config.tensorboard_dir}") + with max_utils.Profiler(config, session_name="flux2klein_inference"): + _, trace_i = pipeline( + prompt=active_prompts, + params=params, + vae_params=vae_params, + qwen3_params=qwen3_params, + vae_bn_mean=vae_bn_mean, + vae_bn_std=vae_bn_std, + transformer_shardings=transformer_shardings, + vae_shardings=vae_shardings, + qwen3_shardings=qwen3_shardings, + height=config.height, + width=config.width, + num_inference_steps=config.num_inference_steps, + batch_size=config.batch_size, + use_latents=use_latents_flag, + latents=latents_to_use, + output_dir=config.output_dir, + output_name=f"rep_{rep+1}_{config.output_name}" if num_reps > 1 else config.output_name, + ) + else: + _, trace_i = pipeline( + prompt=active_prompts, + params=params, + vae_params=vae_params, + qwen3_params=qwen3_params, + vae_bn_mean=vae_bn_mean, + vae_bn_std=vae_bn_std, + transformer_shardings=transformer_shardings, + vae_shardings=vae_shardings, + qwen3_shardings=qwen3_shardings, + height=config.height, + width=config.width, + num_inference_steps=config.num_inference_steps, + batch_size=config.batch_size, + use_latents=use_latents_flag, + latents=latents_to_use, + output_dir=config.output_dir, + output_name=f"rep_{rep+1}_{config.output_name}" if num_reps > 1 else config.output_name, + ) + + tot_time_i = trace_i.get( + "e2e_pipeline_total", + trace_i.get("prompt_encoding", 0.0) + trace_i.get("denoise_loop", 0.0) + trace_i.get("vae_decode", 0.0), + ) + main_traces.append(trace_i) + main_times.append(tot_time_i) + if num_reps > 1: + max_logging.log( + f" -> Rep {rep+1}/{num_reps} Completed: Total={tot_time_i:.4f}s | Qwen3={trace_i.get('qwen3_encoding', 0.0):.4f}s | Denoise={trace_i.get('denoise_loop', 0.0):.4f}s | VAE={trace_i.get('vae_decode', 0.0):.4f}s" + ) + + avg_main_time = sum(main_times) / num_reps + avg_start_to_qwen3 = sum(tr.get("start_to_qwen3", 0.0) for tr in main_traces) / num_reps + avg_prompt_enc = sum(tr.get("qwen3_encoding", tr.get("prompt_encoding", 0.0)) for tr in main_traces) / num_reps + avg_qwen3_to_denoise = sum(tr.get("qwen3_to_denoise", 0.0) for tr in main_traces) / num_reps + avg_denoise = sum(tr.get("denoise_loop", 0.0) for tr in main_traces) / num_reps + avg_denoise_to_vae = sum(tr.get("denoise_to_vae", 0.0) for tr in main_traces) / num_reps + avg_vae_decode = sum(tr.get("vae_decode", 0.0) for tr in main_traces) / num_reps + avg_image_saving = sum(tr.get("image_saving", 0.0) for tr in main_traces) / num_reps + + total_cold_start = load_time + aot_time + warmup_time + max_logging.log("\n" + "=" * 80) - max_logging.log("📊 FLUX.2-KLEIN LATENCY & TIMING BREAKDOWN (PURE MODEL INFERENCE)") + max_logging.log("📊 FLUX.2-KLEIN COMPLETE LATENCY & TIMING BREAKDOWN") max_logging.log("=" * 80) - max_logging.log(f"1) Total Model Loading & Placement Time: {load_time:.2f} seconds ⏱️") - max_logging.log(f"2) Cold-Start / Warmup Pass (XLA Compilation): {warmup_time:.2f} seconds ⏱️") - max_logging.log(f" - Qwen3 Encoding: {warmup_trace.get('prompt_encoding', 0.0):.2f}s") - max_logging.log(f" - Flux Denoising: {warmup_trace.get('denoise_loop', 0.0):.2f}s") - max_logging.log(f" - VAE Decoding: {warmup_trace.get('vae_decode', 0.0):.2f}s") - max_logging.log(f"3) Main Warmed-Up Pass (Pure Model Inference): {main_time:.2f} seconds ⏱️") - max_logging.log(f" - Qwen3 Encoding: {main_trace.get('prompt_encoding', 0.0):.2f}s") - max_logging.log(f" - Flux Denoising: {main_trace.get('denoise_loop', 0.0):.2f}s") - max_logging.log(f" - VAE Decoding: {main_trace.get('vae_decode', 0.0):.2f}s") + max_logging.log(f"1) Model Loading & Placement Time: {load_time:.4f} seconds ⏱️") + max_logging.log(f"2) Concurrent AOT XLA Compilation Time: {aot_time:.4f} seconds ⚡") + max_logging.log(f"3) Warmup Pass Execution Time: {warmup_time:.4f} seconds ⏱️") + max_logging.log(f" - Qwen3 Encoding: {warmup_trace.get('prompt_encoding', 0.0):.4f}s") + max_logging.log(f" - Flux Denoising: {warmup_trace.get('denoise_loop', 0.0):.4f}s") + max_logging.log(f" - VAE Decoding: {warmup_trace.get('vae_decode', 0.0):.4f}s") + max_logging.log(f"👉 TOTAL COLD-START TIME (Loading + AOT + Warmup): {total_cold_start:.4f} seconds 🎯") + rep_label = f" (Average across {num_reps} reps)" if num_reps > 1 else "" + max_logging.log(f"4) Main Warmed-Up Pass (Pure Inference Latency){rep_label}: {avg_main_time:.4f} seconds ⏱️") + max_logging.log(f" - 1. Start -> Qwen3: {avg_start_to_qwen3*1000:.2f} ms ({avg_start_to_qwen3:.4f}s)") + max_logging.log(f" - 2. Qwen3 Encoding: {avg_prompt_enc*1000:.2f} ms ({avg_prompt_enc:.4f}s)") + max_logging.log(f" - 3. Qwen3 -> Denoising: {avg_qwen3_to_denoise*1000:.2f} ms ({avg_qwen3_to_denoise:.4f}s)") + max_logging.log(f" - 4. Flux Denoising Loop: {avg_denoise*1000:.2f} ms ({avg_denoise:.4f}s)") + max_logging.log(f" - 5. Denoising -> VAE: {avg_denoise_to_vae*1000:.2f} ms ({avg_denoise_to_vae:.4f}s)") + max_logging.log(f" - 6. VAE Decoding: {avg_vae_decode*1000:.2f} ms ({avg_vae_decode:.4f}s)") + max_logging.log(f" - 7. Image Saving: {avg_image_saving*1000:.2f} ms ({avg_image_saving:.4f}s)") + max_logging.log(f" - 👉 TOTAL E2E PIPELINE: {avg_main_time*1000:.2f} ms ({avg_main_time:.4f}s)") max_logging.log("=" * 80) max_logging.log("\n=======================================================") diff --git a/src/maxdiffusion/max_utils.py b/src/maxdiffusion/max_utils.py index 37027c27d..608df4448 100644 --- a/src/maxdiffusion/max_utils.py +++ b/src/maxdiffusion/max_utils.py @@ -898,15 +898,23 @@ def initialize_jax_for_gpu(): def maybe_initialize_jax_distributed_system(raw_keys): - if raw_keys["skip_jax_distributed_system"]: + if raw_keys.get("skip_jax_distributed_system", False): max_logging.log("Skipping jax distributed system due to skip_jax_distributed_system=True flag.") return + from jax._src.xla_bridge import backends_are_initialized + + if backends_are_initialized(): + max_logging.log("XLA backends already initialized; skipping jax.distributed.initialize().") + return if is_gpu_backend(raw_keys): max_logging.log("Attempting to initialize the jax distributed system for GPU backend...") initialize_jax_for_gpu() max_logging.log("Jax distributed system initialized on GPU!") else: - jax.distributed.initialize() + try: + jax.distributed.initialize() + except Exception as e: + max_logging.log(f"Warning: jax.distributed.initialize() skipped or failed: {e}") def safe_getattr(obj: Any, name: str, default: Any) -> Any: diff --git a/src/maxdiffusion/models/attention_flax.py b/src/maxdiffusion/models/attention_flax.py index 8b84ae057..4473a1132 100644 --- a/src/maxdiffusion/models/attention_flax.py +++ b/src/maxdiffusion/models/attention_flax.py @@ -2854,6 +2854,8 @@ class FlaxFluxAttention(nn.Module): qkv_bias: bool = False use_base2_exp: bool = False use_experimental_scheduler: bool = False + ulysses_shards: int = -1 + ulysses_attention_chunks: int = 1 def setup(self): if self.attention_kernel in {"flash", "cudnn_flash_te"} and self.mesh is None: @@ -2875,6 +2877,8 @@ def setup(self): float32_qk_product=False, use_base2_exp=self.use_base2_exp, use_experimental_scheduler=self.use_experimental_scheduler, + ulysses_shards=self.ulysses_shards, + ulysses_attention_chunks=self.ulysses_attention_chunks, ) kernel_axes = ("embed", "heads") diff --git a/src/maxdiffusion/models/flux/transformers/transformer_flux_flax.py b/src/maxdiffusion/models/flux/transformers/transformer_flux_flax.py index af8e3763a..8c869dcdd 100644 --- a/src/maxdiffusion/models/flux/transformers/transformer_flux_flax.py +++ b/src/maxdiffusion/models/flux/transformers/transformer_flux_flax.py @@ -132,6 +132,8 @@ class FluxSingleTransformerBlock(nn.Module): precision: jax.lax.Precision = None use_base2_exp: bool = False use_experimental_scheduler: bool = False + ulysses_shards: int = -1 + ulysses_attention_chunks: int = 1 def setup(self): self.mlp_hidden_dim = int(self.dim * self.mlp_ratio) @@ -168,6 +170,8 @@ def setup(self): flash_block_sizes=self.flash_block_sizes, use_base2_exp=self.use_base2_exp, use_experimental_scheduler=self.use_experimental_scheduler, + ulysses_shards=self.ulysses_shards, + ulysses_attention_chunks=self.ulysses_attention_chunks, ) def __call__(self, hidden_states, temb, image_rotary_emb=None): @@ -242,6 +246,8 @@ class FluxTransformerBlock(nn.Module): attention_kernel: str = "dot_product" use_base2_exp: bool = False use_experimental_scheduler: bool = False + ulysses_shards: int = -1 + ulysses_attention_chunks: int = 1 def setup(self): # These contain the parameter projections ("lin"), optimize them using your updated AdaLayerNorm class @@ -260,6 +266,8 @@ def setup(self): flash_block_sizes=self.flash_block_sizes, use_base2_exp=self.use_base2_exp, use_experimental_scheduler=self.use_experimental_scheduler, + ulysses_shards=self.ulysses_shards, + ulysses_attention_chunks=self.ulysses_attention_chunks, ) # REMOVED: self.img_norm2 and self.txt_norm2 completely to stop HBM memory spilling. @@ -767,6 +775,8 @@ class Flux2KleinSingleTransformerBlock(nn.Module): precision: float = None use_global_modulation: bool = False use_swiglu: bool = True + ulysses_shards: int = -1 + ulysses_attention_chunks: int = 1 def setup(self): mlp_hidden_dim = int(self.dim * self.mlp_ratio) @@ -814,6 +824,8 @@ def setup(self): attention_kernel=self.attention_kernel, mesh=self.mesh, flash_block_sizes=self.flash_block_sizes, + ulysses_shards=self.ulysses_shards, + ulysses_attention_chunks=self.ulysses_attention_chunks, ) def __call__(self, hidden_states, temb=None, image_rotary_emb=None, temb_mod=None): @@ -884,6 +896,8 @@ class Flux2KleinTransformerBlock(nn.Module): mlp_ratio: float = 4.0 qkv_bias: bool = True use_global_modulation: bool = False + ulysses_shards: int = -1 + ulysses_attention_chunks: int = 1 def setup(self): if self.use_global_modulation: @@ -924,6 +938,8 @@ def setup(self): weights_dtype=self.weights_dtype, precision=self.precision, qkv_bias=self.qkv_bias, + ulysses_shards=self.ulysses_shards, + ulysses_attention_chunks=self.ulysses_attention_chunks, ) self.ff = FlaxSwiGluFeedForward( self.dim, @@ -1039,6 +1055,8 @@ class Flux2KleinTransformer2DModel(nn.Module, FlaxModelMixin, ConfigMixin): dtype: jnp.dtype = jnp.float32 weights_dtype: jnp.dtype = jnp.float32 precision: float = None + ulysses_shards: int = -1 + ulysses_attention_chunks: int = 1 def setup(self): self.inner_dim = self.num_attention_heads * self.attention_head_dim @@ -1109,6 +1127,8 @@ def setup(self): mlp_ratio=self.mlp_ratio, qkv_bias=self.qkv_bias, use_global_modulation=self.use_global_modulation, + ulysses_shards=self.ulysses_shards, + ulysses_attention_chunks=self.ulysses_attention_chunks, ) double_blocks.append(double_block) self.double_blocks = double_blocks @@ -1128,6 +1148,8 @@ def setup(self): precision=self.precision, mlp_ratio=self.mlp_ratio, use_global_modulation=self.use_global_modulation, + ulysses_shards=self.ulysses_shards, + ulysses_attention_chunks=self.ulysses_attention_chunks, ) single_blocks.append(single_block) self.single_blocks = single_blocks diff --git a/src/maxdiffusion/models/flux/util.py b/src/maxdiffusion/models/flux/util.py index 952519776..f7567e8cb 100644 --- a/src/maxdiffusion/models/flux/util.py +++ b/src/maxdiffusion/models/flux/util.py @@ -300,17 +300,17 @@ def unpack_latents(latents, batch_size, num_channels_latents, height, width): Unpacks packed sequence of shape (batch_size, (height//16)*(width//16), channels*4) back to the unpacked spatial grid shape (batch_size, channels, height//8, width//8). """ - import numpy as np + import jax.numpy as jnp h_latent = height // 8 w_latent = width // 8 # 1. Reshape to split spatial grid and packed channel blocks - latents = np.reshape(latents, (batch_size, h_latent // 2, w_latent // 2, num_channels_latents, 2, 2)) + latents = jnp.reshape(latents, (batch_size, h_latent // 2, w_latent // 2, num_channels_latents, 2, 2)) # 2. Permute dimensions back to unpacked order - latents = np.transpose(latents, (0, 3, 1, 4, 2, 5)) + latents = jnp.transpose(latents, (0, 3, 1, 4, 2, 5)) # 3. Flatten back to 4D unpacked latent shape - latents = np.reshape(latents, (batch_size, num_channels_latents, h_latent, w_latent)) + latents = jnp.reshape(latents, (batch_size, num_channels_latents, h_latent, w_latent)) return latents @@ -398,11 +398,12 @@ def cast_dict_to_bfloat16_inplace(d, device=None, exclude_keywords=None, parent_ is_excluded = exclude_keywords and any(kw.lower() in current_key.lower() for kw in exclude_keywords) target_dtype = jnp.float32 if is_excluded else jnp.bfloat16 - d[k] = v.astype(target_dtype) - if hasattr(d[k], "block_until_ready"): - d[k].block_until_ready() - del v - gc.collect() + if v.dtype != target_dtype: + d[k] = v.astype(target_dtype) + if hasattr(d[k], "block_until_ready"): + d[k].block_until_ready() + del v + gc.collect() # ----------------------------------------------------------------------------- @@ -410,7 +411,9 @@ def cast_dict_to_bfloat16_inplace(d, device=None, exclude_keywords=None, parent_ # ----------------------------------------------------------------------------- -def load_and_convert_flux_klein_weights(safetensors_path, params, num_double_layers, num_single_layers): +def load_and_convert_flux_klein_weights( + safetensors_path, params, num_double_layers, num_single_layers, dtype=None, pt_state_dict=None +): """ Loads weights from safetensors via zero-copy safetensors.numpy and converts them to JAX parameter dictionary. Supports dynamic layer counts (double and single stream blocks) and sharded safetensors directories. @@ -422,28 +425,30 @@ def load_and_convert_flux_klein_weights(safetensors_path, params, num_double_lay import os import gc - pt_state_dict = {} - if os.path.isdir(safetensors_path): - shards = glob.glob(os.path.join(safetensors_path, "*.safetensors")) - max_logging.log(f"Loading sharded weights from directory: {safetensors_path} (Found {len(shards)} shards)...") - for shard in sorted(shards): - max_logging.log(f"Loading shard: {shard}...") - pt_state_dict.update(load_file(shard)) - else: - max_logging.log(f"Loading weights from: {safetensors_path}") - pt_state_dict = load_file(safetensors_path) + if pt_state_dict is None: + pt_state_dict = {} + if os.path.isdir(safetensors_path): + shards = glob.glob(os.path.join(safetensors_path, "*.safetensors")) + max_logging.log(f"Loading sharded weights from directory: {safetensors_path} (Found {len(shards)} shards)...") + for shard in sorted(shards): + max_logging.log(f"Loading shard: {shard}...") + pt_state_dict.update(load_file(shard)) + else: + max_logging.log(f"Loading weights from: {safetensors_path}") + pt_state_dict = load_file(safetensors_path) max_logging.log("Mapping weights to JAX parameters...") expected_pytree = jax.tree_util.tree_map(lambda leaf: leaf, params) first_leaf = jax.tree_util.tree_leaves(params)[0] - target_dtype = first_leaf.dtype + target_dtype = dtype if dtype is not None else first_leaf.dtype - def convert_and_transpose_tensor(tensor, transpose=False): + def convert_and_transpose_tensor(tensor, transpose=False, is_norm=False): if transpose and len(tensor.shape) == 2: tensor = tensor.T - return jnp.array(tensor, dtype=target_dtype) + leaf_dtype = jnp.float32 if is_norm else target_dtype + return jnp.array(tensor, dtype=leaf_dtype) # Global layers params["context_embedder"]["kernel"] = convert_and_transpose_tensor( @@ -562,21 +567,28 @@ def convert_and_transpose_tensor(tensor, transpose=False): return params -def load_and_convert_vae_weights(safetensors_path, jax_params): +def load_and_convert_vae_weights(safetensors_path, jax_params, dtype=None, pt_state_dict=None): """Loads VAE weights from safetensors via zero-copy safetensors.numpy, maps them to JAX, and extracts BN stats.""" from safetensors.numpy import load_file import flax import jax.numpy as jnp - max_logging.log(f"Loading VAE weights from: {safetensors_path}") - pt_state_dict = load_file(safetensors_path) - - def get_pytorch_weight_tensor(key): - return pt_state_dict[key] + if pt_state_dict is None: + max_logging.log(f"Loading VAE weights from: {safetensors_path}") + pt_state_dict = load_file(safetensors_path) # Unfreeze JAX params so we can load the weights jax_params = flax.core.unfreeze(jax_params) + first_leaf = jax.tree_util.tree_leaves(jax_params)[0] + target_dtype = dtype if dtype is not None else first_leaf.dtype + + def get_pytorch_weight_tensor(key, dtype_val=target_dtype): + tensor = pt_state_dict[key] + is_norm = any(kw in key.lower() for kw in ("norm", "layernorm", "rmsnorm", "groupnorm")) + leaf_dtype = jnp.float32 if is_norm else dtype_val + return jnp.array(tensor, dtype=leaf_dtype) + # Map weights max_logging.log("Mapping VAE decoder weights to JAX parameters...") diff --git a/src/maxdiffusion/models/qwen3_flax.py b/src/maxdiffusion/models/qwen3_flax.py index b3ca43003..351c77f7e 100644 --- a/src/maxdiffusion/models/qwen3_flax.py +++ b/src/maxdiffusion/models/qwen3_flax.py @@ -15,7 +15,7 @@ """ import math -from typing import Any, List, Optional, Tuple +from typing import Any, Dict, List, Optional, Tuple from flax import nnx import flax.linen as nn import jax @@ -41,6 +41,12 @@ def __init__( rope_theta: float = 1000000.0, max_position_embeddings: int = 40960, dtype=jnp.float32, + attention_kernel: str = "dot_product", + flash_block_sizes: Optional[Dict[str, int]] = None, + mesh: Optional[jax.sharding.Mesh] = None, + ulysses_shards: int = -1, + ulysses_attention_chunks: int = 1, + max_layer_to_run: Optional[int] = 27, ): self.vocab_size = vocab_size self.hidden_size = hidden_size @@ -53,6 +59,12 @@ def __init__( self.rope_theta = rope_theta self.max_position_embeddings = max_position_embeddings self.dtype = dtype + self.attention_kernel = attention_kernel + self.flash_block_sizes = flash_block_sizes + self.mesh = mesh + self.ulysses_shards = ulysses_shards + self.ulysses_attention_chunks = ulysses_attention_chunks + self.max_layer_to_run = max_layer_to_run # ----------------------------------------------------------------------------- @@ -154,6 +166,9 @@ def rotate_half(x): return q_rot, k_rot +from maxdiffusion.models.attention_flax import AttentionOp + + # ----------------------------------------------------------------------------- # Self Attention (Grouped Query Attention) # ----------------------------------------------------------------------------- @@ -247,33 +262,51 @@ def __call__( k = jnp.repeat(k, gqa_ratio, axis=-2) v = jnp.repeat(v, gqa_ratio, axis=-2) - # 5. Transpose to (batch, num_heads, seq_len, head_dim) for attention - q = jnp.transpose(q, (0, 2, 1, 3)) - k = jnp.transpose(k, (0, 2, 1, 3)) - v = jnp.transpose(v, (0, 2, 1, 3)) - - # 6. Compute attention logits in float32 - q_f = q.astype(jnp.float32) - k_f = k.astype(jnp.float32) - v_f = v.astype(jnp.float32) - - scores = jnp.matmul(q_f, jnp.transpose(k_f, (0, 1, 3, 2))) / math.sqrt(self.config.head_dim) - - # 7. Apply causal attention mask - causal_mask = jnp.tril(jnp.ones((seq_len, seq_len), dtype=jnp.bool_)) - scores = jnp.where(causal_mask, scores, -1e4) - - # 8. Apply padding attention mask if provided - if attention_mask is not None: - p_mask = attention_mask[:, jnp.newaxis, jnp.newaxis, :].astype(jnp.bool_) - scores = jnp.where(p_mask, scores, -1e4) - - # 9. Softmax & Weighted Sum in float32 - probs = jax.nn.softmax(scores, axis=-1) - out = jnp.matmul(probs, v_f).astype(self.config.dtype) + if self.config.attention_kernel != "dot_product": + scale = self.config.head_dim**-0.5 + q_3d = q.reshape((batch_size, seq_len, self.config.num_attention_heads * self.config.head_dim)) + k_3d = k.reshape((batch_size, seq_len, self.config.num_attention_heads * self.config.head_dim)) + v_3d = v.reshape((batch_size, seq_len, self.config.num_attention_heads * self.config.head_dim)) + attn_op = AttentionOp( + mesh=self.config.mesh, + attention_kernel=self.config.attention_kernel, + scale=scale, + heads=self.config.num_attention_heads, + dim_head=self.config.head_dim, + flash_min_seq_length=128, + flash_block_sizes=self.config.flash_block_sizes, + dtype=self.config.dtype, + ulysses_shards=self.config.ulysses_shards, + ulysses_attention_chunks=self.config.ulysses_attention_chunks, + ) + out = attn_op.apply_attention(q_3d, k_3d, v_3d, attention_mask=attention_mask) + else: + # 5. Transpose to (batch, num_heads, seq_len, head_dim) for attention + q = jnp.transpose(q, (0, 2, 1, 3)) + k = jnp.transpose(k, (0, 2, 1, 3)) + v = jnp.transpose(v, (0, 2, 1, 3)) + + # 6. Compute attention logits in float32 + q_f = q.astype(jnp.float32) + k_f = k.astype(jnp.float32) + v_f = v.astype(jnp.float32) + + scores = jnp.matmul(q_f, jnp.transpose(k_f, (0, 1, 3, 2))) / math.sqrt(self.config.head_dim) + + # 7. Apply causal attention mask + causal_mask = jnp.tril(jnp.ones((seq_len, seq_len), dtype=jnp.bool_)) + scores = jnp.where(causal_mask, scores, -1e4) + + # 8. Apply padding attention mask if provided + if attention_mask is not None: + p_mask = attention_mask[:, jnp.newaxis, jnp.newaxis, :].astype(jnp.bool_) + scores = jnp.where(p_mask, scores, -1e4) + + # 9. Softmax & Weighted Sum in float32 + probs = jax.nn.softmax(scores, axis=-1) + out = jnp.matmul(probs, v_f).astype(self.config.dtype) + out = jnp.transpose(out, (0, 2, 1, 3)).reshape((batch_size, seq_len, -1)) - # 10. Reshape back and project out: (batch, seq_len, hidden_size) - out = jnp.transpose(out, (0, 2, 1, 3)).reshape((batch_size, seq_len, -1)) return o_proj(out) @@ -380,7 +413,12 @@ def __call__( ) # 3. Stacked Decoder Layers - for i in range(self.config.num_hidden_layers): + num_layers_to_exec = ( + self.config.num_hidden_layers + if self.config.max_layer_to_run is None + else min(self.config.num_hidden_layers, self.config.max_layer_to_run + 1) + ) + for i in range(num_layers_to_exec): layer = FlaxQwen3DecoderLayer( config=self.config, name=f"layers_{i}", diff --git a/src/maxdiffusion/models/resnet_flax.py b/src/maxdiffusion/models/resnet_flax.py index 79ddcb30e..8371a4432 100644 --- a/src/maxdiffusion/models/resnet_flax.py +++ b/src/maxdiffusion/models/resnet_flax.py @@ -57,9 +57,8 @@ def setup(self): @nn.compact def __call__(self, hidden_states): batch, height, width, channels = hidden_states.shape - hidden_states = jax.image.resize( - hidden_states, shape=(batch, height * 2, width * 2, channels), method="nearest", precision=self.precision - ) + hidden_states = jnp.broadcast_to(hidden_states[:, :, None, :, None, :], (batch, height, 2, width, 2, channels)) + hidden_states = jnp.reshape(hidden_states, (batch, height * 2, width * 2, channels)) hidden_states = nn.with_logical_constraint(hidden_states, ("conv_batch", "height", "keep_2", "out_channels")) diff --git a/src/maxdiffusion/models/vae_flax.py b/src/maxdiffusion/models/vae_flax.py index 72adcbe79..af13327bf 100644 --- a/src/maxdiffusion/models/vae_flax.py +++ b/src/maxdiffusion/models/vae_flax.py @@ -87,11 +87,8 @@ def setup(self): def __call__(self, hidden_states): batch, height, width, channels = hidden_states.shape - hidden_states = jax.image.resize( - hidden_states, - shape=(batch, height * 2, width * 2, channels), - method="nearest", - ) + hidden_states = jnp.broadcast_to(hidden_states[:, :, None, :, None, :], (batch, height, 2, width, 2, channels)) + hidden_states = jnp.reshape(hidden_states, (batch, height * 2, width * 2, channels)) hidden_states = self.conv(hidden_states) return hidden_states diff --git a/src/maxdiffusion/pipelines/flux/flux2klein_pipeline.py b/src/maxdiffusion/pipelines/flux/flux2klein_pipeline.py index 634ec8d9e..dd3a6a101 100644 --- a/src/maxdiffusion/pipelines/flux/flux2klein_pipeline.py +++ b/src/maxdiffusion/pipelines/flux/flux2klein_pipeline.py @@ -31,13 +31,12 @@ from maxdiffusion.max_utils import device_put_replicated from ..pipeline_flax_utils import FlaxDiffusionPipeline from ...models.flux.transformers.transformer_flux_flax import Flux2KleinTransformer2DModel -from ...models.vae_flax import FlaxAutoencoderKL +from ...models.vae_flax import FlaxAutoencoderKL, FlaxDecoderOutput from ...models.qwen3_flax import FlaxQwen3Model from ...schedulers.scheduling_flow_match_flax import FlaxFlowMatchScheduler, compute_empirical_mu from ...models.flux.util import ( pack_latents, - unpack_latents, prepare_latent_image_ids, prepare_text_ids, ) @@ -69,7 +68,35 @@ def __init__( scheduler=scheduler, ) self._config = config + max_layer = getattr(config, "text_encoder_max_layer", 27) + if max_layer is not None and max_layer < 27: + raise ValueError( + f"Invalid configuration `text_encoder_max_layer={max_layer}`. " + f"FLUX.2-Klein requires extracting intermediate prompt embeddings from Qwen3 layers 9, 18, and 27, " + f"so `text_encoder_max_layer` must be >= 27." + ) self.mesh = mesh + self.tokenizer = tokenizer + if self.tokenizer is None: + tokenizer_path = getattr(config, "tokenizer_model_name_or_path", None) or getattr( + config, "pretrained_model_name_or_path", "" + ) + hf_home = os.environ.get("HF_HOME", os.path.expanduser("~/.cache/huggingface")) + repo_cache = os.path.join( + hf_home, + "hub", + f"models--{getattr(config, 'pretrained_model_name_or_path', '').replace('/', '--')}", + "snapshots", + ) + if os.path.exists(repo_cache) and os.listdir(repo_cache): + tokenizer_path = os.path.join(repo_cache, os.listdir(repo_cache)[0]) + + from transformers import Qwen2TokenizerFast + + try: + self.tokenizer = Qwen2TokenizerFast.from_pretrained(tokenizer_path, local_files_only=True) + except Exception: + self.tokenizer = Qwen2TokenizerFast.from_pretrained(tokenizer_path, subfolder="tokenizer", local_files_only=True) # JIT compilation cache self._jitted_qwen3_forward = None @@ -82,7 +109,13 @@ def _setup_jit_functions(self): @jax.jit def qwen3_forward(q_params, ids, mask): - return self.text_encoder.apply({"params": q_params}, input_ids=ids, attention_mask=mask) + _, all_hidden_states = self.text_encoder.apply({"params": q_params}, input_ids=ids, attention_mask=mask) + h_9 = all_hidden_states[9] + h_18 = all_hidden_states[18] + h_27 = all_hidden_states[27] + out = jnp.stack([h_9, h_18, h_27], axis=1) + prompt_embeds = jnp.transpose(out, (0, 2, 1, 3)).reshape((ids.shape[0], ids.shape[1], -1)) + return prompt_embeds @jax.jit def transformer_step(t_params, latents, img_ids, prompt_embeds, txt_ids, vec, timestep, guidance): @@ -97,14 +130,149 @@ def transformer_step(t_params, latents, img_ids, prompt_embeds, txt_ids, vec, ti guidance=guidance, ) + @jax.jit(static_argnums=(4, 5), donate_argnums=(1,)) + def vae_decode(v_params, latents_packed, vae_bn_mean, vae_bn_std, height, width): + batch_size_val = latents_packed.shape[0] + h_latent = height // 8 + w_latent = width // 8 + + vae_bn_mean_seq = vae_bn_mean.reshape(1, 1, 128) + vae_bn_std_seq = vae_bn_std.reshape(1, 1, 128) + + latents_bn = latents_packed * vae_bn_std_seq + vae_bn_mean_seq + latents_unpacked = jnp.reshape(latents_bn, (batch_size_val, h_latent // 2, w_latent // 2, 32, 2, 2)) + latents_unpacked = jnp.transpose(latents_unpacked, (0, 3, 1, 4, 2, 5)) + latents_unpacked = jnp.reshape(latents_unpacked, (batch_size_val, 32, h_latent, w_latent)) + + res = self.vae.apply({"params": v_params}, latents=latents_unpacked, method=self.vae.decode) + return FlaxDecoderOutput(sample=res.sample) + @jax.jit - def vae_decode(v_params, latents_unpatched): - return self.vae.apply({"params": v_params}, latents=latents_unpatched, method=self.vae.decode) + def fused_denoise_loop(t_params, latents, img_ids, prompt_embeds, txt_ids, vec, timesteps, sigmas, guidance): + sigmas_padded = jnp.concatenate([sigmas, jnp.array([0.0], dtype=sigmas.dtype)]) + + def scan_body(cur_latents, step_idx): + t_val = timesteps[step_idx] + t_vec = jnp.broadcast_to(t_val / 1000.0, (cur_latents.shape[0],)) + model_output = self.transformer.apply( + {"params": t_params}, + hidden_states=cur_latents, + img_ids=img_ids, + encoder_hidden_states=prompt_embeds, + txt_ids=txt_ids, + pooled_projections=vec, + timestep=t_vec, + guidance=guidance, + ) + sigma = sigmas_padded[step_idx] + sigma_next = sigmas_padded[step_idx + 1] + prev_sample = cur_latents + model_output.sample * (sigma_next - sigma) + return prev_sample, None + + steps = jnp.arange(timesteps.shape[0]) + final_latents, _ = jax.lax.scan(scan_body, latents, steps) + return final_latents self._jitted_qwen3_forward = qwen3_forward self._jitted_transformer_step = transformer_step + self._jitted_fused_denoise_loop = fused_denoise_loop self._jitted_vae_decode = vae_decode + def _get_dynamic_batch_sharding(self): + """Dynamically infers the batch dimension sharding specification from self.mesh.""" + batch_axes = [axis for axis in ("data", "fsdp") if axis in self.mesh.axis_names and self.mesh.shape[axis] > 1] + spec = P(tuple(batch_axes)) if batch_axes else P(None) + return jax.sharding.NamedSharding(self.mesh, spec) + + def compile_aot_async( + self, params, vae_params, qwen3_params, vae_bn_mean, vae_bn_std, batch_size=1, height=1024, width=1024 + ): + """Triggers AOT compilation for Qwen3, Flux Transformer, and VAE concurrently using ThreadPoolExecutor.""" + self._setup_jit_functions() + max_logging.log("🚀 Pre-compiling XLA graphs for Qwen3, Flux Transformer, and VAE concurrently...") + from concurrent.futures import ThreadPoolExecutor + + seq_len_img = (height // 16) * (width // 16) + seq_len_txt = self._config.max_sequence_length + + dummy_ids = jnp.zeros((batch_size, seq_len_txt), dtype=jnp.int32) + dummy_mask = jnp.ones((batch_size, seq_len_txt), dtype=jnp.int32) + + dummy_latents = jnp.zeros((batch_size, seq_len_img, 128), dtype=jnp.float32) + dummy_img_ids = jnp.zeros((batch_size, seq_len_img, 4), dtype=jnp.int32) + dummy_prompt_embeds = jnp.zeros((batch_size, seq_len_txt, self.transformer.joint_attention_dim), dtype=jnp.bfloat16) + dummy_txt_ids = jnp.zeros((batch_size, seq_len_txt, 4), dtype=jnp.float32) + dummy_t_vec = jnp.zeros((batch_size,), dtype=jnp.float32) + + dummy_bn_mean = jnp.array(vae_bn_mean, dtype=jnp.float32) + dummy_bn_std = jnp.array(vae_bn_std, dtype=jnp.float32) + + data_sharding = self._get_dynamic_batch_sharding() + replicated_sharding = jax.sharding.NamedSharding(self.mesh, P()) + context_sharding = jax.sharding.NamedSharding(self.mesh, P(None, "context")) + + def put_data_on_devices(x, sharding): + if isinstance(x, jax.Array) and hasattr(x, "sharding") and not x.sharding.is_fully_addressable: + return x + if hasattr(sharding, "is_fully_addressable") and sharding.is_fully_addressable: + return jax.device_put(x, sharding) + return device_put_replicated(x, sharding) + + dummy_ids = put_data_on_devices(dummy_ids, data_sharding) + dummy_mask = put_data_on_devices(dummy_mask, data_sharding) + dummy_latents = put_data_on_devices(dummy_latents, data_sharding) + dummy_img_ids = put_data_on_devices(dummy_img_ids, data_sharding) + dummy_prompt_embeds = put_data_on_devices(dummy_prompt_embeds, context_sharding) + dummy_txt_ids = put_data_on_devices(dummy_txt_ids, data_sharding) + dummy_t_vec = put_data_on_devices(dummy_t_vec, data_sharding) + dummy_bn_mean = put_data_on_devices(dummy_bn_mean, replicated_sharding) + dummy_bn_std = put_data_on_devices(dummy_bn_std, replicated_sharding) + + def compile_qwen3(): + t0 = time.perf_counter() + with self.mesh, nn_partitioning.axis_rules(self._config.logical_axis_rules): + self._jitted_qwen3_forward.lower(qwen3_params, dummy_ids, dummy_mask).compile() + max_logging.log(f" -> [AOT COMPILED] Qwen3 Text Encoder in {time.perf_counter() - t0:.2f}s") + + num_steps = getattr(self._config, "num_inference_steps", 4) + dummy_timesteps = put_data_on_devices(jnp.zeros((num_steps,), dtype=jnp.float32), replicated_sharding) + dummy_sigmas = put_data_on_devices(jnp.zeros((num_steps + 1,), dtype=jnp.float32), replicated_sharding) + + def compile_transformer(): + t0 = time.perf_counter() + with self.mesh, nn_partitioning.axis_rules(self._config.logical_axis_rules): + self._jitted_fused_denoise_loop.lower( + params, + dummy_latents, + dummy_img_ids, + dummy_prompt_embeds, + dummy_txt_ids, + None, + dummy_timesteps, + dummy_sigmas, + None, + ).compile() + max_logging.log(f" -> [AOT COMPILED] Fused Flux Transformer Denoise Scan in {time.perf_counter() - t0:.2f}s") + + def compile_vae(): + t0 = time.perf_counter() + with self.mesh, nn_partitioning.axis_rules(self._config.logical_axis_rules): + self._jitted_vae_decode.lower(vae_params, dummy_latents, dummy_bn_mean, dummy_bn_std, height, width).compile() + max_logging.log(f" -> [AOT COMPILED] VAE Decoder in {time.perf_counter() - t0:.2f}s") + + t_start = time.perf_counter() + with ThreadPoolExecutor(max_workers=3) as executor: + futures = [ + executor.submit(compile_qwen3), + executor.submit(compile_transformer), + executor.submit(compile_vae), + ] + for future in futures: + future.result() + aot_duration = time.perf_counter() - t_start + max_logging.log(f"⚡ [AOT CONCURRENT COMPILATION COMPLETE] Total AOT compile time: {aot_duration:.2f}s") + return aot_duration + def _prepare_latents(self, config, batch_size, height, width): num_channels_latents = 32 latent_height = height // 8 @@ -147,8 +315,10 @@ def __call__( use_latents: bool = False, latents: Optional[Any] = None, measure_time: bool = False, + warmup: bool = False, output_dir: str = "output/", output_name: str = "flux2klein_generated_image.png", + profile_target: Optional[str] = None, ): # 1. Setup JIT functions self._setup_jit_functions() @@ -192,6 +362,7 @@ def __call__( sigmas=sigmas_custom, ) + t_pipeline_start = time.perf_counter() trace = {} with self.mesh, nn_partitioning.axis_rules(self._config.logical_axis_rules): @@ -199,60 +370,65 @@ def __call__( proc_cnt = jax.process_count() host_prefix = f"[HOST {proc_id}/{proc_cnt}] " + # Shard pipeline batch inputs across data axis ("data") for SPMD multi-host execution + data_sharding = jax.sharding.NamedSharding(self.mesh, P("data")) + + def put_data_on_devices(x, sharding): + if isinstance(x, jax.Array) and hasattr(x, "sharding") and not x.sharding.is_fully_addressable: + return x + if hasattr(sharding, "is_fully_addressable") and sharding.is_fully_addressable: + return jax.device_put(x, sharding) + return device_put_replicated(x, sharding) + + t0_qwen3_start = time.perf_counter() + trace["start_to_qwen3"] = t0_qwen3_start - t_pipeline_start + max_logging.log(f" -> [TIMING] Start to Qwen3: {trace['start_to_qwen3']:.4f} seconds ⏱️") + # --------------------------------------------------------------------- # PHASE A: Encode Prompt (Qwen3) # --------------------------------------------------------------------- - print(f"{host_prefix} [PHASE A] Encoding {len(prompts)} prompt(s) using JAX Qwen3 on TPU...", flush=True) - t0 = time.perf_counter() - - try: - # Resolve tokenizer path from config - tokenizer_path = self._config.tokenizer_model_name_or_path - hf_home = os.environ.get("HF_HOME", os.path.expanduser("~/.cache/huggingface")) - repo_cache = os.path.join( - hf_home, "hub", f"models--{self._config.pretrained_model_name_or_path.replace('/', '--')}", "snapshots" - ) - if os.path.exists(repo_cache) and os.listdir(repo_cache): - tokenizer_path = os.path.join(repo_cache, os.listdir(repo_cache)[0]) + if not prompts: + raise ValueError("Prompt must be provided to FlaxFlux2KleinPipeline") + if isinstance(prompts, str): + prompts = [prompts] - from transformers import Qwen2TokenizerFast - - try: - tokenizer = Qwen2TokenizerFast.from_pretrained(tokenizer_path, local_files_only=True) - except Exception: - tokenizer = Qwen2TokenizerFast.from_pretrained(tokenizer_path, subfolder="tokenizer", local_files_only=True) + max_logging.log(f"{host_prefix} [PHASE A] Encoding {len(prompts)} prompt(s) using JAX Qwen3 on TPU...") + try: # Tokenize using deterministic explicit template string (version-agnostic across transformers versions) templated_texts = [ f"<|im_start|>user\n{p}<|im_end|>\n<|im_start|>assistant\n\n\n\n\n" for p in prompts ] - inputs = tokenizer( + inputs = self.tokenizer( templated_texts, return_tensors="np", padding="max_length", truncation=True, max_length=seq_len_txt ) prompt_ids = jnp.array(inputs["input_ids"]) prompt_mask = jnp.array(inputs["attention_mask"]) - # Run Text Encoding - hidden_states, all_hidden_states = self._jitted_qwen3_forward(qwen3_params, prompt_ids, prompt_mask) - - # Stack layers 9, 18, 27 to form prompt embeddings - h_9 = all_hidden_states[9] - h_18 = all_hidden_states[18] - h_27 = all_hidden_states[27] - out = jnp.stack([h_9, h_18, h_27], axis=1) - # Transpose shape to [B, seq_len, 3*hidden_size] - prompt_embeds_jax = jnp.transpose(out, (0, 2, 1, 3)).reshape((batch_size, seq_len_txt, -1)) + # Run Text Encoding with sharded input arrays matching compile_aot_async + prompt_ids = put_data_on_devices(prompt_ids, data_sharding) + prompt_mask = put_data_on_devices(prompt_mask, data_sharding) + do_prof_qwen3 = profile_target in ("all", "qwen3") + if do_prof_qwen3: + tb_dir = getattr(self._config, "tensorboard_dir", "/tmp") + jax.profiler.start_trace(os.path.join(tb_dir, "profile_qwen3")) + with jax.named_scope("qwen3_text_encoder"): + prompt_embeds_jax = self._jitted_qwen3_forward(qwen3_params, prompt_ids, prompt_mask) prompt_embeds_jax.block_until_ready() + if do_prof_qwen3: + jax.profiler.stop_trace() except Exception as e: - print(f"❌ {host_prefix} EXCEPTION IN PHASE A (QWEN3 ENCODING): {e}", flush=True) + max_logging.log(f"❌ {host_prefix} EXCEPTION IN PHASE A (QWEN3 ENCODING): {e}") import traceback traceback.print_exc() sys.stdout.flush() raise e - trace["prompt_encoding"] = time.perf_counter() - t0 - max_logging.log(f" -> [TIMING] Prompt Encoding (Qwen3): {trace['prompt_encoding']:.4f} seconds ⏱️") + t0_qwen3_end = time.perf_counter() + trace["qwen3_encoding"] = t0_qwen3_end - t0_qwen3_start + trace["prompt_encoding"] = trace["qwen3_encoding"] + max_logging.log(f" -> [TIMING] Prompt Encoding (Qwen3): {trace['qwen3_encoding']:.4f} seconds ⏱️") proc_id = jax.process_index() proc_cnt = jax.process_count() @@ -260,69 +436,65 @@ def __call__( # Stage Sync 1: Phase A Complete multihost_utils.sync_global_devices("phase_a_complete") - print(f"{host_prefix} Passed Phase A Sync Barrier (phase_a_complete) successfully! ✅", flush=True) - - # Shard pipeline batch inputs across data axis ("data") for SPMD multi-host execution - data_sharding = jax.sharding.NamedSharding(self.mesh, P("data")) - - def put_data_on_devices(x, sharding): - if isinstance(x, jax.Array) and hasattr(x, "sharding") and not x.sharding.is_fully_addressable: - return x - if hasattr(sharding, "is_fully_addressable") and sharding.is_fully_addressable: - return jax.device_put(x, sharding) - return device_put_replicated(x, sharding) + max_logging.log(f"{host_prefix} Passed Phase A Sync Barrier (phase_a_complete) successfully! ✅") latents_jax = put_data_on_devices(latents_jax, data_sharding) - prompt_embeds_jax = put_data_on_devices(prompt_embeds_jax, data_sharding) txt_ids_val = put_data_on_devices(txt_ids_val, data_sharding) img_ids_val = put_data_on_devices(img_ids_val, data_sharding) - print( + max_logging.log( f"{host_prefix} DIAGNOSTIC TENSORS BEFORE PHASE B:\n" f" latents_jax: shape={latents_jax.shape}, dtype={latents_jax.dtype}, sharding={getattr(latents_jax, 'sharding', None)}\n" f" prompt_embeds_jax: shape={prompt_embeds_jax.shape}, dtype={prompt_embeds_jax.dtype}, sharding={getattr(prompt_embeds_jax, 'sharding', None)}\n" f" txt_ids_val: shape={txt_ids_val.shape}, dtype={txt_ids_val.dtype}, sharding={getattr(txt_ids_val, 'sharding', None)}\n" - f" img_ids_val: shape={img_ids_val.shape}, dtype={img_ids_val.dtype}, sharding={getattr(img_ids_val, 'sharding', None)}", - flush=True, + f" img_ids_val: shape={img_ids_val.shape}, dtype={img_ids_val.dtype}, sharding={getattr(img_ids_val, 'sharding', None)}" ) # Stage Sync 2: Pre-Phase B Start multihost_utils.sync_global_devices("pre_phase_b_start") - print(f"{host_prefix} Passed Pre-Phase B Sync Barrier (pre_phase_b_start) successfully! ✅", flush=True) + max_logging.log(f"{host_prefix} Passed Pre-Phase B Sync Barrier (pre_phase_b_start) successfully! ✅") + + t0_denoise_start = time.perf_counter() + trace["qwen3_to_denoise"] = t0_denoise_start - t0_qwen3_end + max_logging.log(f" -> [TIMING] Qwen3 to Denoising Overhead: {trace['qwen3_to_denoise']:.4f} seconds ⏱️") # --------------------------------------------------------------------- # PHASE B: Denoising Loop (Flux Transformer - Standalone Step JIT) # --------------------------------------------------------------------- - print( - f"{host_prefix} [PHASE B] Running {num_inference_steps}-step E2E Denoising Loop on a batch of {batch_size} images...", - flush=True, + steps_to_run = num_inference_steps + max_logging.log( + f"{host_prefix} [PHASE B] Running fused {steps_to_run}-step E2E Denoising Loop Scan on a batch of {batch_size} images (warmup={warmup})..." ) - t0 = time.perf_counter() try: guidance_vec_val = None vec_val = None - - for step_idx in range(num_inference_steps): - timestep = scheduler_state.timesteps[step_idx] - t_vec = jnp.full((batch_size,), timestep / 1000.0, dtype=latents_jax.dtype) - - model_output = self._jitted_transformer_step( - params, latents_jax, img_ids_val, prompt_embeds_jax, txt_ids_val, vec_val, t_vec, guidance_vec_val - ) - - prev_sample, _ = self.scheduler.step( - state=scheduler_state, - model_output=model_output.sample, - timestep=scheduler_state.timesteps[step_idx], - sample=latents_jax, - return_dict=False, + replicated_sharding = jax.sharding.NamedSharding(self.mesh, P()) + timesteps_device = put_data_on_devices(scheduler_state.timesteps, replicated_sharding) + sigmas_device = put_data_on_devices(scheduler_state.sigmas, replicated_sharding) + + do_prof_denoise = profile_target in ("all", "denoise") + if do_prof_denoise: + tb_dir = getattr(self._config, "tensorboard_dir", "/tmp") + jax.profiler.start_trace(os.path.join(tb_dir, "profile_denoise")) + with jax.named_scope("fused_flux_denoise_loop"): + latents_jax = self._jitted_fused_denoise_loop( + params, + latents_jax, + img_ids_val, + prompt_embeds_jax, + txt_ids_val, + vec_val, + timesteps_device, + sigmas_device, + guidance_vec_val, ) - latents_jax = prev_sample + latents_jax.block_until_ready() + if do_prof_denoise: + jax.profiler.stop_trace() - latents_jax.block_until_ready() except Exception as e: - print(f"❌ {host_prefix} EXCEPTION IN DENOISE LOOP: {e}", flush=True) + max_logging.log(f"❌ {host_prefix} EXCEPTION IN DENOISE LOOP: {e}") import traceback traceback.print_exc() @@ -331,32 +503,41 @@ def put_data_on_devices(x, sharding): # Stage Sync 3: Phase B Complete multihost_utils.sync_global_devices("phase_b_complete") - print(f"{host_prefix} Passed Phase B Sync Barrier (phase_b_complete) successfully! ✅", flush=True) + max_logging.log(f"{host_prefix} Passed Phase B Sync Barrier (phase_b_complete) successfully! ✅") - trace["denoise_loop"] = time.perf_counter() - t0 + t0_denoise_end = time.perf_counter() + trace["denoise_loop"] = t0_denoise_end - t0_denoise_start max_logging.log(f" -> [TIMING] Denoising Loop (Flux): {trace['denoise_loop']:.4f} seconds ⏱️") # --------------------------------------------------------------------- # PHASE C: Decode Latents (VAE Decoder) # --------------------------------------------------------------------- max_logging.log("[PHASE C] Decoding final latents to RGB image using JAX VAE decoder on TPU...") - t0 = time.perf_counter() - - # Apply Channel-wise Batch Normalization Scaling in packed sequence format (denormalize) - vae_bn_mean_seq = vae_bn_mean.reshape(1, 1, 128) - vae_bn_std_seq = vae_bn_std.reshape(1, 1, 128) - latents_bn = latents_jax * vae_bn_std_seq + vae_bn_mean_seq - # Unpack packed latents back to spatial grid - latents_unpacked = unpack_latents(latents_bn, batch_size, 32, height, width) - - # Decode VAE latents to RGB pixels - decoded_out = self._jitted_vae_decode(vae_params, latents_unpacked) - # VAE output is in decoded_out.sample + # Decode VAE latents to RGB pixels using fused JIT vae_decode + data_sharding = self._get_dynamic_batch_sharding() + replicated_sharding = jax.sharding.NamedSharding(self.mesh, P()) + latents_jax = put_data_on_devices(latents_jax, data_sharding) + vae_bn_mean_jax = put_data_on_devices(jnp.array(vae_bn_mean, dtype=jnp.float32), replicated_sharding) + vae_bn_std_jax = put_data_on_devices(jnp.array(vae_bn_std, dtype=jnp.float32), replicated_sharding) + + t0_vae_start = time.perf_counter() + trace["denoise_to_vae"] = t0_vae_start - t0_denoise_end + max_logging.log(f" -> [TIMING] Denoising to VAE Overhead: {trace['denoise_to_vae']:.4f} seconds ⏱️") + + do_prof_vae = profile_target in ("all", "vae") + if do_prof_vae: + tb_dir = getattr(self._config, "tensorboard_dir", "/tmp") + jax.profiler.start_trace(os.path.join(tb_dir, "profile_vae")) + with jax.named_scope("vae_decoder"): + decoded_out = self._jitted_vae_decode(vae_params, latents_jax, vae_bn_mean_jax, vae_bn_std_jax, height, width) images_rgb = decoded_out.sample images_rgb.block_until_ready() + if do_prof_vae: + jax.profiler.stop_trace() - trace["vae_decode"] = time.perf_counter() - t0 + t0_vae_end = time.perf_counter() + trace["vae_decode"] = t0_vae_end - t0_vae_start max_logging.log(f" -> [TIMING] VAE Decoding: {trace['vae_decode']:.4f} seconds ⏱️") # --------------------------------------------------------------------- @@ -364,15 +545,15 @@ def put_data_on_devices(x, sharding): # --------------------------------------------------------------------- max_logging.log("Postprocessing and saving generated images...") saved_paths = [] - # Clamp pixels and scale to [0, 255] - images_rgb = jnp.clip((images_rgb + 1.0) / 2.0, 0.0, 1.0) + # Perform pixel scaling, clamping, and uint8 conversion directly on TPU hardware + images_uint8 = jnp.clip((images_rgb + 1.0) * 127.5, 0.0, 255.0).astype(jnp.uint8) if jax.process_count() > 1: - images_numpy = multihost_utils.process_allgather(images_rgb, tiled=True) + images_numpy = multihost_utils.process_allgather(images_uint8, tiled=True) else: - images_numpy = np.array(images_rgb) + images_numpy = np.array(images_uint8) for b_idx in range(batch_size): - image_np = np.array(images_numpy[b_idx] * 255.0, dtype=np.uint8) + image_np = np.array(images_numpy[b_idx]) # Transpose channel dimension if shape is (C, H, W) instead of (H, W, C) if image_np.shape[0] == 3: image_np = image_np.transpose(1, 2, 0) @@ -386,8 +567,15 @@ def put_data_on_devices(x, sharding): batch_output_name = output_name output_png_path = os.path.join(output_dir, batch_output_name) - img.save(output_png_path) + img.save(output_png_path, format="PNG", compress_level=1) max_logging.log(f" -> Saved image: {output_png_path} | Prompt: '{prompts[b_idx]}'") saved_paths.append(output_png_path) + t0_save_end = time.perf_counter() + trace["image_saving"] = t0_save_end - t0_vae_end + trace["e2e_pipeline_total"] = t0_save_end - t_pipeline_start + + max_logging.log(f" -> [TIMING] Image Saving: {trace['image_saving']:.4f} seconds ⏱️") + max_logging.log(f" -> [TIMING] E2E Pipeline Total: {trace['e2e_pipeline_total']:.4f} seconds ⏱️") + return saved_paths, trace diff --git a/src/maxdiffusion/schedulers/scheduling_flow_match_flax.py b/src/maxdiffusion/schedulers/scheduling_flow_match_flax.py index 8e9f38ff4..f649b08da 100644 --- a/src/maxdiffusion/schedulers/scheduling_flow_match_flax.py +++ b/src/maxdiffusion/schedulers/scheduling_flow_match_flax.py @@ -300,6 +300,7 @@ def step( sample: jnp.ndarray, to_final: bool = False, return_dict: bool = True, + step_index: Optional[int] = None, ) -> Union[FlaxFlowMatchSchedulerOutput, Tuple]: """ Propagates the sample with the flow matching scheduler. @@ -317,12 +318,17 @@ def step( Whether this is the final step. return_dict (`bool`): Whether to return a `FlaxFlowMatchSchedulerOutput` object. + step_index (`Optional[int]`): + Optional direct step index to bypass dynamic _find_timestep_id calculation. Returns: `FlaxFlowMatchSchedulerOutput` or `tuple`: A tuple (`prev_sample`, `state`) or a `FlaxFlowMatchSchedulerOutput` object containing the previous sample and the updated state. """ - timestep_id = self._find_timestep_id(state, timestep) + if step_index is not None: + timestep_id = step_index + else: + timestep_id = self._find_timestep_id(state, timestep) sigma = state.sigmas[timestep_id] def get_next_sigma(): diff --git a/src/maxdiffusion/tests/generate_flux2klein_smoke_test.py b/src/maxdiffusion/tests/generate_flux2klein_smoke_test.py index 24362d35d..cad425439 100644 --- a/src/maxdiffusion/tests/generate_flux2klein_smoke_test.py +++ b/src/maxdiffusion/tests/generate_flux2klein_smoke_test.py @@ -17,6 +17,7 @@ import os import unittest import pytest +import jax import numpy as np from PIL import Image @@ -54,22 +55,23 @@ def test_flux2klein_4b_smoke(self): "run_name=smoke_test_4b", f"output_dir={output_dir}", "jax_cache_dir=/tmp/cache_dir", - "skip_jax_distributed_system=True", f"prompt={PROMPT}", "height=512", "width=512", - "batch_size=1", + f"per_device_batch_size={1.0 / jax.device_count()}", "seed=42", - "ici_fsdp_parallelism=-1", "weights_dtype=bfloat16", "activations_dtype=bfloat16", "precision=DEFAULT", + "num_reps=5", ] generate_flux2klein.main(args) - self.assertTrue(os.path.exists(out_path), "Smoke test 4B failed to produce output image!") - test_image = np.array(Image.open(out_path)).astype(np.uint8) + rep_out_path = os.path.join(output_dir, "rep_1_flux2klein_generated_image.png") + final_out_path = rep_out_path if os.path.exists(rep_out_path) else out_path + self.assertTrue(os.path.exists(final_out_path), "Smoke test 4B failed to produce output image!") + test_image = np.array(Image.open(final_out_path)).astype(np.uint8) self.assertEqual(base_image.shape, test_image.shape) ssim_compare = ssim(base_image, test_image, channel_axis=-1, data_range=255) @@ -97,27 +99,28 @@ def test_flux2klein_9b_smoke(self): "run_name=smoke_test_9b", f"output_dir={output_dir}", "jax_cache_dir=/tmp/cache_dir", - "skip_jax_distributed_system=True", f"prompt={PROMPT}", "height=512", "width=512", - "batch_size=1", + f"per_device_batch_size={1.0 / jax.device_count()}", "seed=42", - "ici_fsdp_parallelism=-1", "weights_dtype=bfloat16", "activations_dtype=bfloat16", "precision=DEFAULT", + "num_reps=5", ] generate_flux2klein.main(args) - self.assertTrue(os.path.exists(out_path), "Smoke test 9B failed to produce output image!") - test_image = np.array(Image.open(out_path)).astype(np.uint8) + rep_out_path = os.path.join(output_dir, "rep_1_flux2klein_generated_image.png") + final_out_path = rep_out_path if os.path.exists(rep_out_path) else out_path + self.assertTrue(os.path.exists(final_out_path), "Smoke test 9B failed to produce output image!") + test_image = np.array(Image.open(final_out_path)).astype(np.uint8) self.assertEqual(base_image.shape, test_image.shape) ssim_compare = ssim(base_image, test_image, channel_axis=-1, data_range=255) print(f"\n[SMOKE TEST 9B] SSIM Score: {ssim_compare:.6f}") - self.assertGreaterEqual(ssim_compare, 0.80) + self.assertGreaterEqual(ssim_compare, 0.8) if __name__ == "__main__": diff --git a/src/maxdiffusion/tests/images/ref_flux2klein_9b.png b/src/maxdiffusion/tests/images/ref_flux2klein_9b.png index 594464a8f..c27f959ac 100644 Binary files a/src/maxdiffusion/tests/images/ref_flux2klein_9b.png and b/src/maxdiffusion/tests/images/ref_flux2klein_9b.png differ