"""Isolated full-Trainer A1/A2 runtime and fail-closed checkpoint contract. A1: genuinely aliased frozen prefix plus independent tail-eight, optionally penultimate causal cross. A2: one body AV pass; detached KV/text-base references feed private QA LoRA. No loss, data, optimizer or checkpoint cadence is changed. """ from __future__ import annotations from contextlib import contextmanager from dataclasses import asdict, dataclass from functools import wraps import hashlib from functools import lru_cache import json import os from pathlib import Path import re import sys from types import MethodType import torch from .shared_prefix import SharedPrefixReplay, install_on_swa from .shared_av import install_shared_av from .coupled_tail_cross import CoupledTailCrossRuntime, register_coupled_tail_cross ARCH_KEY = "isolated_architecture" FORMAT = "hfd_a1_c_production_v1" @dataclass(frozen=True) class ArchitectureSpec: arm: str interaction: str = "none" prefix_layers: int = 28 cross_layer: int = 34 bridge_dim: int = 256 heads: int = 8 max_cross_tokens: int = 32 bridge_seed: int = 941 control_token_arm: str = "C" def __post_init__(self): if self.arm not in {"a1", "a2"} or self.interaction not in {"none", "cross"}: raise ValueError("Only a1/a2 and none/cross are implemented") if self.arm != "a1" or self.interaction != "cross" or self.control_token_arm not in {"A", "B", "C"}: raise ValueError("Control A/B uses the identical A1 cross backbone") if self.arm == "a2" and self.interaction != "none": raise ValueError("Selected A2 is shared-AV stop-gradient without cross") if self.prefix_layers < 1: raise ValueError("A frozen nonempty prefix is required") if self.interaction == "cross" and not self.prefix_layers <= self.cross_layer < self.prefix_layers + 7: raise ValueError("Cross must leave a later AV KV layer in the private tail") def code_identity(): # Hash architecture implementation only: launcher/metric edits do not # change checkpoint semantics. No dependence on an experiment workspace. root = Path(__file__).parent names = ("runtime.py", "control_tokens.py", "shared_prefix.py", "shared_av.py", "coupled_tail_cross.py", "coupled_joint.py") return {"files": {name: hashlib.sha256((root / name).read_bytes()).hexdigest() for name in names}} def architecture_contract(spec): return dict(format=FORMAT, spec=asdict(spec), implementation_sha256=code_identity(), frozen_prefix_parameter_alias=spec.arm == "a1", private_mutable_qa_kv=True, answer_text_never_writes_av=True, qa_to_body_gradient=False, semantics=("shared_frozen_prefix_independent_tail8" if spec.arm == "a1" else "single_body_av_detached_snapshot_private_text_lora_live_detached_body_base")) class PrefixCrossRuntime(CoupledTailCrossRuntime): """Body-first rendezvous: record prefix BEFORE QA attempts prefix replay. The original paired executor is QA-first, which cannot compose with true prefix compute sharing. Here the body pauses at cross-layer entry; QA replays the already computed prefix and rendezvouses at the same layer. Both native self-attentions/crosses finish, then QA suffix, then body suffix. No foreign history is fabricated, and the formal later QA call consumes its real result. """ def enable_with_prefix(self, prefix, *, optimizer=None, prepared=False): model = self._model_ref() self.prefix = SharedPrefixReplay(model.vlm_model.model, model.qa_model.vlm_model.model, prefix_layers=prefix, history_mode="lineage", alias_prefix=True) try: super().enable(optimizer=optimizer, prepared=prepared) original_condition = model._encode_qwen_stream_input @wraps(original_condition) def paired(_model, **kwargs): with self.prefix.pair(): return original_condition(**kwargs) self._patch(model, "_encode_qwen_stream_input", MethodType(paired, model)) except Exception: if self.enabled: super().close() self.prefix.close() raise return self @property def stats(self): return {**super().stats, "frozen_prefix_replay": dict(self.prefix.stats) if hasattr(self, "prefix") else {}, "paired_execution_order": "body_prefix_then_qa_replay_then_paired_tail"} def _body_forward(self, *args, **kwargs): pair = self._pair if pair is None or pair.phase == "consumed": return self._original_body_forward(*args, **kwargs) if pair.phase != "body_entry" or args: raise RuntimeError("Unexpected shared-prefix body model entry") self._validate_model_args(kwargs) pair.body_args, pair.phase = kwargs, "body_running" with self.prefix._packet.body(): result = self._original_body_forward(**kwargs) if pair.phase != "qa_suffix" or pair.qa_encoded is None: raise RuntimeError("Body-first rendezvous did not finish the real QA branch") pair.phase = "body_return" return result def _body_layer(self, hidden, **kwargs): pair = self._pair if pair is None or pair.phase == "consumed": self._counts["text_layer_bypasses"] += 1 return self._original_body_layer(hidden, **kwargs) if pair.phase != "body_running": raise RuntimeError("Body-first cross layer is out of order") pair.body_layer_input, pair.body_layer_kwargs = hidden, kwargs packet = self.prefix._packet if packet.role != "body": raise RuntimeError("Shared prefix body scope was lost") # The prefix is complete although the body suffix has not returned. # This is needed for a sound replay lineage certificate, not just speed. packet.body_cache = kwargs["past_key_values"] pair.phase = "qa_prefix" model = self._model_ref() model.qa_model.vlm_processor = model.vlm_processor packet.role = None # Suspend the body scope only while QA executes. try: with packet.qa(), torch.set_grad_enabled(pair.qa_grad): pair.qa_encoded = self._original_encode( model.qa_model.vlm_model, pair.qa_cache, pair.packets, pair.active, memory_seconds=model.qa_model.memory_seconds, context_limit=model.qa_model.av_context_limit, prepared=pair.prepared) finally: packet.role = "body" if pair.phase != "qa_suffix" or not hasattr(pair, "body_layer_output"): raise RuntimeError("QA did not finish the shared-prefix cross rendezvous") self._counts["qa_av_forwards"] += 1 return pair.body_layer_output def _qa_layer(self, hidden, **kwargs): pair = self._pair if pair is None or pair.phase == "consumed": self._counts["text_layer_bypasses"] += 1 return self._original_qa_layer(hidden, **kwargs) if pair.phase != "qa_prefix" or not hasattr(pair, "body_layer_input"): raise RuntimeError("QA has no same-packet body cross input") pair.qa_layer_input, pair.qa_layer_kwargs = hidden, kwargs body, qa = self._joint_layers(pair.body_layer_input, pair.body_layer_kwargs, pair) pair.body_layer_output = body pair.phase = "qa_suffix" return qa def close(self): super().close() if hasattr(self, "prefix"): self.prefix.close() class ArchitectureRuntime: def __init__(self, model, spec): if hasattr(model, "_isolated_architecture_runtime"): raise RuntimeError("Architecture already registered") layers = model.vlm_model.model.layers if len(layers) != spec.prefix_layers + 8: raise ValueError("Architecture manifest must match frozen-prefix plus exactly eight private tail layers") self.model, self.spec = model, spec self.contract = architecture_contract(spec) self.operator = None self.installed = False self._inference_originals = [] if spec.arm == "a1" and spec.interaction == "cross": self.operator = register_coupled_tail_cross(model, layer_index=spec.cross_layer, bridge_dim=spec.bridge_dim, heads=spec.heads, max_cross_tokens=spec.max_cross_tokens, seed=spec.bridge_seed) self.operator.__class__ = PrefixCrossRuntime model._isolated_architecture_runtime = self def install(self, *, optimizer=None, prepared=False): if self.installed or not prepared: raise RuntimeError("Install once after final prepare/device/dtype placement") trainable_before = {id(p) for p in self.model.parameters() if p.requires_grad} optimizer_before = None if optimizer is None else [id(p) for g in optimizer.param_groups for p in g["params"]] if self.spec.arm == "a2": self.operator = install_shared_av(self.model, joint=False, optimizer=optimizer, prepared=True) elif self.spec.interaction == "none": self.operator = install_on_swa(self.model, prefix_layers=self.spec.prefix_layers, history_mode="lineage", alias_prefix=True) else: self.operator.enable_with_prefix(self.spec.prefix_layers, optimizer=optimizer, prepared=True) if {id(p) for p in self.model.parameters() if p.requires_grad} != trainable_before: raise RuntimeError("Architecture install changed trainable parameters after optimizer creation") if optimizer is not None: after = [id(p) for g in optimizer.param_groups for p in g["params"]] if after != optimizer_before or len(after) != len(set(after)) or set(after) != trainable_before: raise RuntimeError("Architecture optimizer ownership mismatch") self.installed = True return self @property def stats(self): return dict(arm=self.spec.arm, interaction=self.spec.interaction, installed=self.installed, runtime={} if self.operator is None else dict(self.operator.stats)) def close(self): for target, name, original in reversed(self._inference_originals): setattr(target, name, original) self._inference_originals.clear() if self.operator is not None: self.operator.close() def serialize_inference_dispatch(self): """Protect single-threaded AV hooks from the native answer worker. Serialize Python forward dispatch only, not the whole speech job. AV and answer KV remain independent; speech keeps its original CUDA stream and records actual queue/PCM availability, including lock contention. """ from threading import RLock if not self.installed or self._inference_originals: raise RuntimeError("Install inference dispatch guard exactly once after architecture install") lock = RLock() for target, name in ((self.model, "_encode_qwen_stream_input"), (self.model.qa_model.vlm_model.model, "forward")): original = getattr(target, name) def guarded(*args, _original=original, **kwargs): with lock: return _original(*args, **kwargs) self._inference_originals.append((target, name, original)) setattr(target, name, wraps(original)(guarded)) def resolve_checkpoint(path): """Match formal Trainer best-or-latest resolution, but never silently reset.""" root = Path(path).expanduser().resolve() if (root / "model.safetensors").is_file(): return root folder = root / "checkpoints" if (folder / "best_ckpt" / "model.safetensors").is_file(): return folder / "best_ckpt" choices = sorted((p for p in folder.glob("ckpt_*") if re.fullmatch(r"ckpt_\d+", p.name)), key=lambda p: int(p.name[5:])) if choices and (choices[-1] / "model.safetensors").is_file(): return choices[-1] raise FileNotFoundError(f"Explicit resume checkpoint not found: {root}") def validate_checkpoint(path, contract, expected_keys=None): from safetensors import safe_open path = Path(path) manifest = json.loads((path / "hfd_fused_delta_manifest.json").read_text()) if manifest.get("format") != "hfd_qwen_serial_delta_v1": raise RuntimeError("Unsupported formal checkpoint format") if manifest.get("contract", {}).get(ARCH_KEY) != contract: raise RuntimeError("Architecture/checkpoint mismatch or unmarked legacy checkpoint; cold migration is not resume") with safe_open(str(path / "model.safetensors"), framework="pt", device="cpu") as handle: keys = set(handle.keys()) if any(name.startswith("qa_model.") and ".base." in name for name in keys): raise RuntimeError("Checkpoint illegally contains an independent QA base") if expected_keys is not None and keys != set(expected_keys): raise RuntimeError("Architecture checkpoint tensor key set does not match the live model") return manifest @contextmanager def trainer_architecture(spec): """Patch only lifecycle boundaries; execute the entire unchanged Trainer.""" from psi.trainers.posttrain_full_duplex import PosttrainFullDuplexTrainer as Trainer originals = {name: getattr(Trainer, name) for name in ( "init_models", "prepare", "_build_hfdv2_checkpoint_contract", "resume_from_checkpoint", "log", "evaluate")} runtimes = [] def initialize(self): originals["init_models"](self) if int(getattr(self.train_cfg, "stage", 0)) != 0: raise ValueError("Architecture long training is stage 0 only") from .control_tokens import install_token_rows install_token_rows(self.model, arm=spec.control_token_arm) self._architecture = ArchitectureRuntime(self.model, spec) runtimes.append(self._architecture) def contract(self, *args, **kwargs): result = dict(originals["_build_hfdv2_checkpoint_contract"](self, *args, **kwargs)) result[ARCH_KEY] = self._architecture.contract return result def prepare(self, accelerator): if self.train_cfg.data_parallel in {"fsdp", "deepspeed"}: raise RuntimeError("A1/C requires DDP") result = originals["prepare"](self, accelerator) model = accelerator.unwrap_model(self.model) if model is not self._architecture.model: raise RuntimeError("Unexpected model replacement during prepare") self._architecture.install(optimizer=self.optimizer, prepared=True) from .control_tokens import enable_packet_write enable_packet_write(model) if accelerator.is_main_process: project = Path(self.project_dir) project.mkdir(parents=True, exist_ok=True) (project / "architecture_manifest.json").write_text(json.dumps(self._architecture.contract, indent=2)) (project / "control_token_contract.json").write_text(json.dumps(model.control_token_contract, indent=2)) model.vlm_processor.tokenizer.save_pretrained(project / "control_tokenizer") parameters = {id(p): p for p in model.parameters()} (project / "architecture_parameters.json").write_text(json.dumps(dict( unique_parameters=sum(p.numel() for p in parameters.values()), actual_trainable_parameters=sum(p.numel() for p in parameters.values() if p.requires_grad), note="Measured after bridge registration and alias; earlier formal init log predates bridge registration"), indent=2)) for tracker in accelerator.trackers: if tracker.name == "wandb": tracker.run.config.update({ARCH_KEY: self._architecture.contract}, allow_val_change=False) return result def resume(self): path = self.train_cfg.resume_from_checkpoint if path is not None: checkpoint = resolve_checkpoint(path) model = self.accelerator.unwrap_model(self.model) validate_checkpoint(checkpoint, self._architecture.contract, self._hfdv2_delta_keys(model)) return originals["resume_from_checkpoint"](self) def log(self, metrics, start_time=None): model = self.accelerator.unwrap_model(self.model) anchor = getattr(model, "control_token_semantic_anchor", None) if anchor is not None: weight = model.control_condition_embedding.weight.float() target = anchor.float().to(weight.device) metrics = dict(metrics) metrics["train/control_token_anchor_cosine"] = float( torch.nn.functional.cosine_similarity(weight, target, dim=-1).mean().detach().cpu()) metrics["train/control_token_anchor_rmse"] = float( (weight - target).square().mean().sqrt().detach().cpu()) result = originals["log"](self, metrics, start_time=start_time) from .control_metrics import flush extra = flush(self, "train") if extra: self.accelerator.log(extra, step=self.global_step) if self.accelerator.is_main_process: from psi.runtime.local_metrics import append_metrics append_metrics(self.project_dir, self.global_step, extra) return result def evaluate(self): from .control_metrics import flush getattr(self, "_token_control_meters", {}).pop("eval", None) was_training = self.model.training self.model.eval() try: result = originals["evaluate"](self) finally: self.model.train(was_training) extra = flush(self, "eval") if extra: self.accelerator.log(extra, step=self.global_step) if self.accelerator.is_main_process: from psi.runtime.local_metrics import append_metrics append_metrics(self.project_dir, self.global_step, extra) if result is not None: result.update({k.split("/",1)[1]: v for k,v in extra.items()}) return result for name, method in {"init_models": initialize, "prepare": prepare, "_build_hfdv2_checkpoint_contract": contract, "resume_from_checkpoint": resume, "log": log, "evaluate": evaluate}.items(): setattr(Trainer, name, method) try: yield finally: for runtime in reversed(runtimes): runtime.close() for name, method in originals.items(): setattr(Trainer, name, method) def load_inference_model(run_dir, checkpoint_step, launch_config, device, *, expected_spec=None): """Formal fused-delta loader with explicit architecture reconstruction. Standard unadapted inference fails its contract check, rather than silently loading A2 LoRA onto an old independent QA base. No KV is resumed across sources. """ if not str(device).startswith("cuda") or os.environ.get("PSI_INFER_FLOAT32", "0") == "1": raise NotImplementedError("Architecture inference currently targets CUDA BF16 only; CPU/FP32 QA backend parity is unvalidated") from psi.models.full_duplex_swa import FullDuplexSwaModel from psi.data.seed import qwen_streaming_audio as streaming run = Path(run_dir).resolve() name = str(checkpoint_step) name = "best_ckpt" if name in {"best", "best_ckpt"} else name if name.startswith("ckpt_") else "ckpt_" + name path = run / "checkpoints" / name if not path.is_dir(): path = run / name saved = json.loads((path / "hfd_fused_delta_manifest.json").read_text()).get("contract", {}).get(ARCH_KEY) if not isinstance(saved, dict) or saved.get("format") != FORMAT: raise RuntimeError("An architecture-marked checkpoint is required") spec = ArchitectureSpec(**saved["spec"]) if expected_spec is not None and spec != expected_spec: raise RuntimeError("Requested inference architecture does not match checkpoint") current = architecture_contract(spec) validate_checkpoint(path, current) class InferenceModel(FullDuplexSwaModel): def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) from transformers import Qwen2_5OmniProcessor from .control_tokens import install_token_rows self.vlm_processor = Qwen2_5OmniProcessor.from_pretrained(launch_config.model.model_name_or_path) install_token_rows(self, arm=spec.control_token_arm) ArchitectureRuntime(self, spec) # Before formal expected-delta check/load. original_contract = streaming.hfdv2_train_infer_contract def augmented(*args, **kwargs): return {**original_contract(*args, **kwargs), ARCH_KEY: current} streaming.hfdv2_train_infer_contract = augmented try: with torch.inference_mode(False): model = InferenceModel.from_streaming_checkpoint(run, checkpoint_step, launch_config, device) from transformers import AutoTokenizer from .control_tokens import enable_packet_write model.vlm_processor.tokenizer = AutoTokenizer.from_pretrained(run / "control_tokenizer") if tuple(model.vlm_processor.tokenizer.convert_tokens_to_ids(t) for t in ("<|listen|>", "<|motion|>", "<|idle|>")) != model.control_token_ids: raise RuntimeError("Saved inference state token IDs disagree with training") model._isolated_architecture_runtime.install(prepared=True) enable_packet_write(model) model._isolated_architecture_runtime.serialize_inference_dispatch() return model finally: streaming.hfdv2_train_infer_contract = original_contract