"""Process-local single-AV Thinker experiment; never modifies formal sources. Install ONLY after trainer.prepare/device/dtype placement, on the unwrapped model. Existing QA LoRA Parameter objects (and their optimizer state) survive. AV tokens always use body W; only the private text continuation uses W + LoRA. This is deliberate token-routed adaptation, NOT reuse of W+LoRA AV caches. close() restores the global dispatcher and drops pending graphs, but does not restore the discarded independent QA base. Dispose of the experimental model after closing. This prevents retaining a hidden second copy of base weights. """ from __future__ import annotations from dataclasses import replace import weakref import torch from torch import nn from torch.nn import functional as F from transformers.cache_utils import DynamicCache from psi.models import causal_av from psi.models.swa_qa_branch import BranchLoRALinear class ReferencedLinear(nn.Module): """Unregistered, live reference to a body's Linear; no base Parameter copy.""" def __init__(self, source: nn.Linear, *, detach: bool): super().__init__() if not isinstance(source, nn.Linear): raise TypeError("Shared AV requires ordinary body Linear projections") self._source_ref = weakref.ref(source) self.detach = bool(detach) self.in_features = source.in_features self.out_features = source.out_features @property def source(self): source = self._source_ref() if source is None: raise RuntimeError("Shared AV body projection no longer exists") return source @property def weight(self): value = self.source.weight return value.detach() if self.detach else value @property def bias(self): value = self.source.bias return value.detach() if self.detach and value is not None else value def forward(self, x): return F.linear(x, self.weight, self.bias) class ReferencedRMSNorm(nn.Module): """Qwen RMSNorm with the same live-reference and gradient ownership rule.""" def __init__(self, source: nn.Module, *, detach: bool): super().__init__() if not hasattr(source, "variance_epsilon") or not hasattr(source, "weight"): raise TypeError("Shared AV expects Qwen RMSNorm") self._source_ref = weakref.ref(source) self.detach = bool(detach) @property def source(self): source = self._source_ref() if source is None: raise RuntimeError("Shared AV body norm no longer exists") return source @property def weight(self): value = self.source.weight return value.detach() if self.detach else value def forward(self, x): dtype = x.dtype x = x.float() x = x * torch.rsqrt(x.square().mean(-1, keepdim=True) + self.source.variance_epsilon) return self.weight * x.to(dtype) def clone_av_snapshot(cache, *, detach: bool): """Clone BOTH container and storage; QA rebind/compact/write cannot mutate AV.""" def copy(value): if value is None: return None return (value.detach() if detach else value).clone() past = cache.past_key_values if past is not None: past = DynamicCache(ddp_cache_data=[ (copy(layer.keys), copy(layer.values)) for layer in past.layers ]) return replace(cache, past_key_values=past, attention_mask=copy(cache.attention_mask), token_birth_windows=copy(cache.token_birth_windows), qa_cache=None) def _adapter_parameters(qa): return {id(p): p for module in qa.modules() if isinstance(module, BranchLoRALinear) for part in (module.down, module.up) for p in part.parameters()} def audit_optimizer(model, optimizer): """The QA update clock/clipping group must contain adapters and nothing else.""" adapters = _adapter_parameters(model.qa_model) all_ids, qa_ids = [], [] for group in optimizer.param_groups: ids = [id(p) for p in group["params"]] all_ids.extend(ids) if group.get("group_name") == "swa_qa": qa_ids.extend(ids) if len(all_ids) != len(set(all_ids)): raise RuntimeError("Optimizer contains duplicate shared Parameter objects") if set(qa_ids) != set(adapters): raise RuntimeError("swa_qa optimizer group must contain exactly existing QA LoRA parameters") known = {id(p) for p in model.parameters()} if not set(all_ids) <= known: raise RuntimeError("Optimizer retains discarded independent QA base parameters") class SharedAVRuntime: def __init__(self, model, *, joint=False, optimizer=None, prepared=False): if not prepared: raise RuntimeError("Install shared AV only AFTER trainer.prepare; pass prepared=True") if getattr(causal_av, "_shared_av_runtime", None) is not None: raise RuntimeError("A shared AV dispatcher is already installed in this process") body = model.vlm_model qa = model.qa_model if qa is None: raise ValueError("Shared AV experiment requires the existing online QA branch") reader = qa.vlm_model if body is reader: raise ValueError("QA text shell and body must be distinct modules") body_layers, qa_layers = body.model.layers, reader.model.layers if len(body_layers) != len(qa_layers) or len(body_layers) < 8: raise ValueError("Shared AV needs matching Qwen models with at least eight layers") split = len(body_layers) - 8 # Validate completely before releasing/replacing any independent weights. projections = (("self_attn", ("q_proj", "k_proj", "v_proj", "o_proj")), ("mlp", ("gate_proj", "up_proj", "down_proj"))) for index in range(split, len(body_layers)): for owner, names in projections: for name in names: dst = getattr(getattr(qa_layers[index], owner), name) src = getattr(getattr(body_layers[index], owner), name) if not isinstance(dst, BranchLoRALinear) or not isinstance(src, nn.Linear): raise TypeError("Expected body Linear and QA BranchLoRALinear tail projections") if dst.down.weight.shape[1] != src.in_features or dst.up.weight.shape[0] != src.out_features: raise ValueError("Body and QA projection shapes differ") if dst.down.weight.device != src.weight.device: raise RuntimeError("Install after both branches have reached their final device") for name in ("input_layernorm", "post_attention_layernorm"): if not hasattr(getattr(body_layers[index], name), "variance_epsilon"): raise TypeError("Expected body Qwen RMSNorm") frozen = list(body_layers[:split]) + [body.model.embed_tokens, body.lm_head] if any(p.requires_grad for module in frozen for p in module.parameters()): raise RuntimeError("Shared AV requires frozen body prefix, embedding and lm_head") if not hasattr(body.model.norm, "variance_epsilon"): raise TypeError("Expected body final Qwen RMSNorm") if joint and not any(p.requires_grad for layer in body_layers[-8:] for p in layer.parameters()): raise RuntimeError("joint=True requires the body's last eight layers to be trainable") if optimizer is not None: audit_optimizer(model, optimizer) self.joint = bool(joint) self._body_ref = weakref.ref(body) self._reader_ref = weakref.ref(reader) self.pending = None self.body_av_forwards = 0 self.qa_av_reads = 0 self.closed = False self._original_encode = causal_av.encode_av adapter_ids = set(_adapter_parameters(qa)) for index in range(split): qa_layers[index] = body_layers[index] for index in range(split, len(body_layers)): for owner, names in projections: for name in names: dst = getattr(getattr(qa_layers[index], owner), name) src = getattr(getattr(body_layers[index], owner), name) dst.base = ReferencedLinear(src, detach=not joint) for name in ("input_layernorm", "post_attention_layernorm"): setattr(qa_layers[index], name, ReferencedRMSNorm(getattr(body_layers[index], name), detach=not joint)) for name in ("embed_tokens", "rotary_emb"): setattr(reader.model, name, getattr(body.model, name)) # The formal body can train final norm. Never freeze it or directly # alias its trainable weight into QA: isolation must apply here too. reader.model.norm = ReferencedRMSNorm(body.model.norm, detach=not joint) reader.lm_head = body.lm_head if hasattr(reader, "audio_tower"): reader.audio_tower = None if hasattr(reader, "visual"): reader.visual = None # Config is not a parameter; synchronize special ids and text geometry. reader.config = body.config shared_ids = {id(p) for p in body.parameters()} reader_ids = {id(p) for p in reader.parameters()} if reader_ids - shared_ids != adapter_ids: raise RuntimeError("QA retains unrecognized independent non-LoRA parameters") if optimizer is not None: audit_optimizer(model, optimizer) self._dispatch = self.encode_av causal_av.encode_av = self._dispatch causal_av._shared_av_runtime = self @property def stats(self): return {"body_av_forwards": self.body_av_forwards, "qa_av_reads": self.qa_av_reads} def encode_av(self, qwen, cache, packets, active, **kwargs): if self.closed: raise RuntimeError("Shared AV experiment has been closed") body, reader = self._body_ref(), self._reader_ref() prepared = kwargs.get("prepared") if qwen is body: if self.pending is not None: self.pending = None raise RuntimeError("Previous body AV write was not paired with a QA read") if prepared is None: raise RuntimeError("Shared AV pairing requires an explicit PreparedAVBatch") if self.joint and cache.past_key_values is not None: if any(t.requires_grad for layer in cache.past_key_values.layers for t in (layer.keys, layer.values)): raise RuntimeError("Detach body streaming history between windows; joint only covers current AV") result = self._original_encode(qwen, cache, packets, active, **kwargs) self.body_av_forwards += 1 self.pending = (prepared, packets, result) return result if qwen is reader: pending, self.pending = self.pending, None if pending is None: raise RuntimeError("QA AV encode without an immediately paired body write") expected, expected_packets, result = pending if prepared is not expected or packets is not expected_packets: raise RuntimeError("QA AV read must use the identical prepared body packet batch") if cache.source_ids != result[-1].source_ids: raise RuntimeError("QA/body source ids disagree in shared AV pairing") snapshot = clone_av_snapshot(result[-1], detach=not self.joint) hidden = result[0] if self.joint else result[0].detach() self.qa_av_reads += 1 return hidden, result[1], result[2], result[3], snapshot if self.pending is not None: self.pending = None raise RuntimeError("Another AV model interrupted shared AV body/QA pairing") return self._original_encode(qwen, cache, packets, active, **kwargs) def close(self): self.pending = None if self.closed: return if causal_av.encode_av is not self._dispatch: raise RuntimeError("Cannot restore shared AV dispatcher: another patch replaced it") causal_av.encode_av = self._original_encode if getattr(causal_av, "_shared_av_runtime", None) is self: del causal_av._shared_av_runtime self.closed = True def __enter__(self): return self def __exit__(self, *_): self.close() def install_shared_av(model, *, joint=False, optimizer=None, prepared=False): return SharedAVRuntime(model, joint=joint, optimizer=optimizer, prepared=prepared)