Blame for src/celestis_rl/streaming.py

"""Recomputed first-order linear-head VJPs for centered policy scores.

Exact mode removes auxiliary sampling noise; sampled mode preserves the
selected estimator. Neither stores an [N,K,D] autograd activation. Both still
need an exact forward softmax normalizer. These are portable PyTorch kernels,
not fused CUDA, and produce dense parameter gradients. Higher-order AD is
intentionally unsupported. q, correction weights, and temperature are fixed.
"""
from dataclasses import dataclass
import math
import torch
from torch import Tensor
from torch.autograd.function import once_differentiable
from .estimators import canonical_q, Correction
from .losses import _prepare, _reduce


@dataclass
class CenteredScores:
    centered_logits: Tensor
    action_log_probs: Tensor  # detached, only for coefficients


def _check(h,w,a,bias,chunk,temperature):
    if h.ndim!=2 or w.ndim!=2 or h.shape[1]!=w.shape[1] or min(*h.shape,*w.shape)<1:
        raise ValueError('nonempty hidden [N,D] and weight [V,D] required')
    if not h.is_floating_point() or not w.is_floating_point():
        raise ValueError('floating point model tensors required')
    if a.shape!=(len(h),) or a.dtype!=torch.long:
        raise ValueError('actions [N] must be int64')
    if h.device!=w.device or a.device!=h.device or ((a<0)|(a>=len(w))).any():
        raise ValueError('invalid actions/devices')
    if bias is not None and (bias.shape!=(len(w),) or bias.device!=w.device or not bias.is_floating_point()):
        raise ValueError('invalid bias')
    if isinstance(chunk,bool) or not isinstance(chunk,int) or chunk<1:
        raise ValueError('chunk size must be positive integer')
    if not math.isfinite(temperature) or temperature<=0:
        raise ValueError('temperature must be positive finite')
    if not torch.isfinite(h).all() or not torch.isfinite(w).all() or (bias is not None and not torch.isfinite(bias).all()):
        raise ValueError('nonfinite active model tensor')


class _Centered(torch.autograd.Function):
    @staticmethod
    def forward(ctx,h,w,a,bias,q,ids,coeff,chunk,rows,temperature,exact):
        dtype=torch.float64 if any(x is not None and x.dtype==torch.float64 for x in (h,w,bias)) else torch.float32
        ctx.dtype=dtype;ctx.chunk=chunk;ctx.rows=rows;ctx.temperature=temperature;ctx.exact=exact
        ctx.has_bias=bias is not None
        empty=torch.empty(0,device=h.device,dtype=dtype)
        qsum=q.sum(-1,keepdim=True,dtype=torch.float64) if exact else empty
        ctx.save_for_backward(h,w,a,bias if bias is not None else empty,q,ids,coeff,qsum)
        with torch.autocast(device_type=h.device.type,enabled=False):
            x=h.to(dtype)
            za=(x*w[a].to(dtype)).sum(-1)
            if bias is not None:za=za+bias[a].to(dtype)
            za=za/temperature
            normal=torch.full((len(h),),-torch.inf,device=h.device,dtype=dtype)
            expected=torch.zeros_like(normal)
            for start in range(0,len(w),chunk):
                end=min(start+chunk,len(w))
                z=x@w[start:end].to(dtype).T
                if bias is not None:z=z+bias[start:end].to(dtype)
                z=z/temperature
                normal=torch.logaddexp(normal,torch.logsumexp(z,-1))
                if exact:
                    qc=(q[:,start:end].double()/qsum).to(dtype)
                    expected=expected+(qc*z).sum(-1)
            if exact:
                centered=za-expected
            else:
                centered=torch.zeros_like(normal)
                for start in range(0,len(h),rows):
                    end=min(start+rows,len(h));ii=ids[start:end]
                    z=(x[start:end,None,:]*w[ii].to(dtype)).sum(-1)
                    if bias is not None:z=z+bias[ii].to(dtype)
                    centered[start:end]=(z*coeff[start:end].to(dtype)).sum(-1)/temperature
            lp=za-normal
            if not torch.isfinite(centered).all() or not torch.isfinite(lp).all():
                raise FloatingPointError('nonfinite centered linear scores')
        ctx.mark_non_differentiable(lp)
        return centered,lp

    @staticmethod
    @once_differentiable
    def backward(ctx,gcenter,glogp):
        h,w,a,bias,q,ids,coeff,qsum=ctx.saved_tensors
        dtype=ctx.dtype;temp=ctx.temperature
        if gcenter is None:gcenter=torch.zeros(len(h),device=h.device,dtype=dtype)
        with torch.autocast(device_type=h.device.type,enabled=False):
            x=h.to(dtype);g=gcenter.to(dtype)/temp
            gh=torch.zeros_like(x) if ctx.needs_input_grad[0] else None
            gw=torch.zeros(w.shape,device=w.device,dtype=dtype) if ctx.needs_input_grad[1] else None
            gb=torch.zeros(len(w),device=w.device,dtype=dtype) if ctx.has_bias and ctx.needs_input_grad[3] else None
            if ctx.exact:
                for start in range(0,len(w),ctx.chunk):
                    end=min(start+ctx.chunk,len(w))
                    c=-(q[:,start:end].double()/qsum).to(dtype)
                    active=(a>=start)&(a<end)
                    rr=active.nonzero().flatten()
                    c[rr,a[rr]-start]+=1
                    c=c*g[:,None]
                    if gh is not None:gh.add_(c@w[start:end].to(dtype))
                    if gw is not None:gw[start:end]=c.T@x
                    if gb is not None:gb[start:end]=c.sum(0)
            else:
                for start in range(0,len(h),ctx.rows):
                    end=min(start+ctx.rows,len(h));ii=ids[start:end]
                    c=coeff[start:end].to(dtype)*g[start:end,None]
                    if gh is not None:gh[start:end]=(c[:,:,None]*w[ii].to(dtype)).sum(1)
                    if gw is not None:
                        contributions=(c[:,:,None]*x[start:end,None,:]).reshape(-1,w.shape[1])
                        gw.index_add_(0,ii.reshape(-1),contributions)
                    if gb is not None:gb.index_add_(0,ii.reshape(-1),c.reshape(-1))
        return (None if gh is None else gh.to(h.dtype),None if gw is None else gw.to(w.dtype),None,
                None if gb is None else gb.to(bias.dtype),None,None,None,None,None,None,None)


def centered_linear_scores(hidden: Tensor,weight: Tensor,actions: Tensor,*,q: Tensor | None=None,
                           ids: Tensor | None=None,coefficients: Tensor | None=None,
                           bias: Tensor | None=None,chunk_size: int=4096,row_chunk_size: int=64,
                           temperature: float=1.) -> CenteredScores:
    """Compute z(a)-E_q[z], or sum_j coefficients_j*z(ids_j).

    Provide q for exact mode OR ids/coefficients for a sampled zero-sum
    measure (typically one action weight +1 and auxiliary weights summing -1).
    The action-normalization pass remains exact. No gradient is returned for
    the fixed measure, sampler, IDs, temperature, or detached log probabilities.
    """
    _check(hidden,weight,actions,bias,chunk_size,temperature)
    if isinstance(row_chunk_size,bool) or not isinstance(row_chunk_size,int) or row_chunk_size<1:
        raise ValueError('row_chunk_size must be positive integer')
    exact=q is not None
    empty=torch.empty(0,device=hidden.device)
    if exact:
        if ids is not None or coefficients is not None:raise ValueError('choose exact OR sampled measure')
        if q.shape!=(len(hidden),len(weight)) or q.device!=hidden.device or not q.is_floating_point():
            raise ValueError('q shape/device/type mismatch')
        if not torch.isfinite(q).all() or (q<0).any() or not torch.allclose(q.double().sum(-1),torch.ones(len(hidden),device=q.device,dtype=torch.float64),atol=2e-6,rtol=0):
            raise ValueError('invalid actual sampler probabilities')
        q=q.detach();ids=empty.long();coefficients=empty
    else:
        if ids is None or coefficients is None or ids.ndim!=2 or ids.shape!=coefficients.shape or ids.shape[0]!=len(hidden) or ids.shape[1]<1:
            raise ValueError('selected ids and coefficients [N,K] required')
        if ids.dtype!=torch.long or not coefficients.is_floating_point() or ids.device!=hidden.device or coefficients.device!=hidden.device:
            raise ValueError('invalid selected measure devices/types')
        if ((ids<0)|(ids>=len(weight))).any() or not torch.isfinite(coefficients).all():
            raise ValueError('invalid selected measure')
        if (coefficients.double().sum(-1).abs()>2e-6).any():
            raise ValueError('centered measure must have zero total mass')
        q=empty;coefficients=coefficients.detach()
    c,lp=_Centered.apply(hidden,weight,actions,bias,q,ids,coefficients,chunk_size,row_chunk_size,float(temperature),exact)
    return CenteredScores(c,lp)


def linear_token_loss(hidden: Tensor,weight: Tensor,actions: Tensor,q: Tensor,mask: Tensor,
                      returns: Tensor,*,correction: Correction | None=None,bias: Tensor | None=None,
                      beta: float=.1,baseline: Tensor | None=None,outer_weights: Tensor | None=None,
                      normalizer: float | None=None,temperature: float=1.,chunk_size: int=4096,
                      row_chunk_size: int=64) -> tuple[Tensor,dict]:
    """Memory-bounded full or sampled token update with inherited normalization.

    hidden [B,T,D], q [B,T,V]. None correction selects exact streaming.
    Masked hidden/return entries are sanitized BEFORE any arithmetic.
    """
    if hidden.ndim!=3 or actions.shape!=hidden.shape[:2] or q.shape!=(*hidden.shape[:2],len(weight)):
        raise ValueError('hidden [B,T,D], actions [B,T], q [B,T,V] required')
    if actions.dtype!=torch.long or mask.shape!=actions.shape or mask.dtype!=torch.bool or mask.device!=hidden.device:
        raise ValueError('invalid actions/mask')
    _,r=_prepare(hidden[...,0],mask,returns,beta)
    h=hidden.masked_fill(~mask[...,None],0);a=actions.masked_fill(~mask,0)
    if a.device!=h.device:raise ValueError('action device mismatch')
    sampler=canonical_q(q,mask)
    qa=sampler.gather(-1,a[...,None]).squeeze(-1)
    if (qa[mask]<=0).any():raise ValueError('observed action absent from behavior support')
    b,t,d=h.shape
    opts={'bias':bias,'temperature':temperature,'chunk_size':chunk_size,'row_chunk_size':row_chunk_size}
    if correction is None:
        scores=centered_linear_scores(h.reshape(-1,d),weight,a.reshape(-1),q=sampler.reshape(b*t,-1),**opts)
        route='exact_streaming'
    else:
        if not torch.equal(correction.mask,mask):raise ValueError('correction mask mismatch')
        if not torch.isfinite(correction.weights).all() or (correction.weights<0).any():raise ValueError('invalid correction weights')
        sums=correction.weights.sum(-1)
        if not torch.allclose(sums[mask],torch.ones_like(sums[mask]),atol=2e-6,rtol=0):raise ValueError('correction mass mismatch')
        ids=torch.cat((a[...,None],correction.ids),-1)
        coeff=torch.cat((mask[...,None].to(correction.weights.dtype),-correction.weights),-1)
        scores=centered_linear_scores(h.reshape(-1,d),weight,a.reshape(-1),ids=ids.reshape(b*t,-1),coefficients=coeff.reshape(b*t,-1),**opts)
        route='sampled_recomputed'
    lp=scores.action_log_probs.reshape(b,t).masked_fill(~mask,0)
    lq=qa.log().masked_fill(~mask,0)
    coefficient=r-beta*(lp-lq)
    if baseline is not None:
        if baseline.shape!=(b,t) or baseline.device!=h.device:raise ValueError('baseline [B,T] required')
        base=baseline.detach().masked_fill(~mask,0)
        if not torch.isfinite(base).all():raise ValueError('nonfinite active baseline')
        coefficient=coefficient-base
    coefficient=coefficient.detach().masked_fill(~mask,0)
    per=-(coefficient*scores.centered_logits.reshape(b,t)).sum(-1)
    loss=_reduce(per,outer_weights,normalizer)
    if not torch.isfinite(loss):raise FloatingPointError('nonfinite linear loss')
    return loss,{'route':route,'coefficient':coefficient,'per_response_surrogate':per.detach()}