scripts/agentx/alignment.py

# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
"""Offline, run-audited sequence alignment. No fixed receive-time window.

All identities remain inferred. Duration is an alignment feature, so duration
agreement after matching is NOT independent proof of timing accuracy.
"""
import bisect
import math

import numpy as np
import numba
from numba import njit

VERSION = "rank-sequence-alignment-v2"


def correlation(a, b):
    if len(a) < 3 or np.std(a) == 0 or np.std(b) == 0:
        return None
    return float(np.corrcoef(a, b)[0, 1])


def sglang_groups(records, expected, full_timed, ordinals):
    lengths = [len(full_timed[r]) for r in expected]
    if not lengths[0] or len(set(lengths)) != 1:
        raise ValueError("SGLang ordinal join requires equal full-capture timed counts")
    n = lengths[0]
    ref = np.asarray(full_timed[expected[0]])
    checks = {}
    for rank in expected[1:]:
        other = np.asarray(full_timed[rank])
        global_corr = correlation(ref, other)
        windows, failed, minimum = 0, [], 1.0
        for low in range(0, n, 1000):
            high = min(n, low+1000)
            if high-low < 3:
                continue
            zero = correlation(ref[low:high], other[low:high])
            prev = correlation(ref[low+1:high], other[low:high-1])
            following = correlation(ref[low:high-1], other[low+1:high])
            windows += 1
            if zero is None or zero < .7 or any(c is not None and c >= zero for c in (prev, following)):
                failed.append(low)
            if zero is not None:
                minimum = min(minimum, zero)
        if global_corr is None or global_corr < .995 or failed:
            raise ValueError(f"SGLang ordinal audit failed for rank {rank}: {failed}")
        checks[str(rank)] = dict(global_correlation=global_corr, checked_windows=windows,
                                 failed_windows=failed, minimum_zero_lag_correlation=minimum)
    groups = {}
    selected = {}
    for row in records:
        ordinal = ordinals[row["source_line"]]
        groups.setdefault(ordinal, []).append(row)
        selected.setdefault(row["dp_rank"], []).append(ordinal)
    output = []
    for ordinal, group in sorted(groups.items()):
        output.append((sorted(group, key=lambda r: r["dp_rank"]),
                       dict(method="validated_full_capture_timed_ordinal", full_capture_ordinal=ordinal)))
    return output, dict(method="validated_full_capture_timed_ordinal", version=VERSION,
                        full_capture_timed_counts=dict(zip(expected, lengths)),
                        selected_ordinal_ranges={r:[min(v),max(v)] for r,v in selected.items()},
                        zero_lag_vs_adjacent_lag_audits=checks,
                        warning="Run-specific inferred correspondence; not a publisher-certified global ID.")


@njit
def segment(at, aw, bt, bw, gap):
    n, m = len(aw), len(bw)
    cost = np.zeros((n+1, m+1))
    trace = np.zeros((n+1, m+1), dtype=np.uint8)
    for i in range(1,n+1):
        cost[i,0], trace[i,0] = gap*i, 1
    for j in range(1,m+1):
        cost[0,j], trace[0,j] = gap*j, 2
    for i in range(1,n+1):
        for j in range(1,m+1):
            scale = .00003 + .0001*max(aw[i-1],bw[j-1])
            duration = math.log1p(abs(aw[i-1]-bw[j-1])/scale)
            timing = .10*min(abs(at[i-1]-bt[j-1])/(.002+max(aw[i-1],bw[j-1])),10.)
            values = (cost[i-1,j-1]+duration+timing, cost[i-1,j]+gap, cost[i,j-1]+gap)
            op = 0
            if values[1] < values[op]:
                op = 1
            if values[2] < values[op]:
                op = 2
            cost[i,j], trace[i,j] = values[op], op
    result = np.full(n,-1,dtype=np.int64)
    i,j=n,m
    while i or j:
        op=trace[i,j]
        if op==0:
            result[i-1]=j-1
            i-=1
            j-=1
        elif op==1:
            i-=1
        else:
            j-=1
    return result


def candidates(a,b,radius):
    center=np.searchsorted(b[:,0],a[:,0]).clip(0,len(b)-1)
    js=center[:,None]+np.arange(-radius,radius+1)
    valid=(js>=0)&(js<len(b))
    safe=js.clip(0,len(b)-1)
    maximum=np.maximum(a[:,1,None],b[safe,1])
    delta=abs(a[:,1,None]-b[safe,1])
    score=delta/(.00005+.0001*maximum)
    score+=.10*abs(a[:,0,None]-b[safe,0])/(.002+maximum)
    score[~valid]=np.inf
    order=np.argsort(score,axis=1,kind="stable")
    ix=np.arange(len(a))
    best=safe[ix,order[:,0]]
    values=score[ix,order[:,0]]
    margin=score[ix,order[:,1]]-values
    good=(delta[ix,order[:,0]]<.00005+.0001*maximum[ix,order[:,0]])&(values<2)&(margin>.2)
    return best,good


def anchors(a,b,radius):
    best,good=candidates(a,b,radius)
    inverse,other_good=candidates(b,a,radius)
    indices=np.flatnonzero(good&other_good[best]&(inverse[best]==np.arange(len(a))))
    tails,tail_indices=[],[]
    predecessors=np.full(len(indices),-1,dtype=int)
    for k,i in enumerate(indices):
        j=int(best[i]);pos=bisect.bisect_left(tails,j)
        if pos:
            predecessors[k]=tail_indices[pos-1]
        if pos==len(tails):
            tails.append(j);tail_indices.append(k)
        else:
            tails[pos]=j;tail_indices[pos]=k
    selected=[]
    if tail_indices:
        k=tail_indices[-1]
        while k>=0:
            selected.append(indices[k]);k=predecessors[k]
    selected=np.asarray(selected[::-1],dtype=int)
    return selected,best[selected],len(indices)-len(selected)


def pair(a,b,radius=12,gap=4.):
    ia,ib,crossings=anchors(a,b,radius)
    result=np.full(len(a),-1,dtype=int)
    result[ia]=ib
    weak=np.zeros(len(a),dtype=bool)
    longest=(0,0);unresolved=[]
    bounds=list(zip([-1,*ia,len(a)],[-1,*ib,len(b)]))
    for (i,j),(ni,nj) in zip(bounds[:-1],bounds[1:]):
        lowa,lowb=int(i+1),int(j+1);ni,nj=int(ni),int(nj)
        n,m=ni-lowa,nj-lowb
        if not n or not m:
            continue
        if n*m>longest[0]*longest[1]:
            longest=(n,m)
        if n*m>1_000_000:
            unresolved.append((lowa,ni,lowb,nj));continue
        local=segment(a[lowa:ni,0],a[lowa:ni,1],b[lowb:nj,0],b[lowb:nj,1],gap)
        matched=local>=0
        selected=np.arange(lowa,ni)[matched]
        result[selected]=local[matched]+lowb
        # One isolated record among many missing steps is not enough to locate
        # a unique iteration, even if the minimum-cost path picks one.
        if max(n,m)>32 and min(n,m)<4:
            weak[selected]=True
    return result,weak,dict(anchors=len(ia),anchor_crossings_rejected=crossings,
                            largest_segment=longest,unresolved_segments=unresolved,
                            matched=int((result>=0).sum()),weak_sparse_matches=int(weak.sum()))


def vllm_groups(records,expected):
    by_rank={r:[] for r in expected}
    for row in records:
        by_rank[row["dp_rank"]].append(row)
    for values in by_rank.values():
        values.sort(key=lambda r:r["counter_id"])
        if not values:
            raise ValueError("No positive records for an expected vLLM rank")
    origin=min(row["received_at_ns"] for row in records)
    data={r:np.asarray([((x["received_at_ns"]-origin)/1e9,x["wall_time"]) for x in values])
          for r,values in by_rank.items()}
    if any(np.any(np.diff(a[:,0])<0) for a in data.values()):
        raise ValueError("Recorder clock is not monotonic within a rank")
    primary=max(expected,key=lambda r:len(by_rank[r]))
    # Explicit independent reference for this validated capture; not a producer ID.
    secondary=3 if 3 in expected and primary!=3 else next(r for r in expected if r!=primary)
    maps,weakness,summaries={},{},{}
    for ref in (primary,secondary):
        maps[ref]={ref:np.arange(len(data[ref]))}
        weakness[ref]={ref:np.zeros(len(data[ref]),dtype=bool)}
        summaries[str(ref)]={}
        for rank in expected:
            if rank==ref:
                continue
            mapping,weak,summary=pair(data[ref],data[rank])
            maps[ref][rank],weakness[ref][rank]=mapping,weak
            summaries[str(ref)][str(rank)]=summary
    final={primary:maps[primary][primary].copy()}
    consistency={};sensitivity={};rejected={}
    middle=maps[primary][secondary]
    for rank in expected:
        if rank==primary:
            continue
        direct=maps[primary][rank]
        transitive=np.full(len(direct),-1,dtype=int)
        exists=middle>=0
        transitive[exists]=maps[secondary][rank][middle[exists]]
        both=(direct>=0)&(transitive>=0)
        conflict=both&(direct!=transitive)
        alternative,_,_=pair(data[primary],data[rank],radius=24,gap=3.)
        competing=(direct>=0)&(alternative>=0)&(direct!=alternative)
        # A conservative run can abstain on a true prefill duration difference;
        # abstention is not a conflicting positive correspondence.
        accept=~(conflict|competing|weakness[primary][rank])
        result=direct.copy();result[~accept]=-1
        final[rank]=result
        consistency[str(rank)]=dict(shared_pairs=int(both.sum()),
                                    matching_pairs=int((both&~conflict).sum()),
                                    competing_pairs=int(conflict.sum()))
        sensitivity[str(rank)]=dict(competing_positive_pairs=int(competing.sum()),
                                    conservative_abstentions=int(((direct>=0)&(alternative<0)).sum()))
        rejected[str(rank)]=int(((direct>=0)&~accept).sum())
    used=set();output=[]
    for i in range(len(data[primary])):
        members=[]
        for rank in expected:
            j=int(final[rank][i])
            if j<0:
                continue
            row=by_rank[rank][j]
            if row["source_line"] in used:
                raise ValueError("Sequence matching reused an observation")
            used.add(row["source_line"]);members.append(row)
        output.append((members,dict(method="monotone_anchor_sequence", reference_rank=primary,
                                   reference_counter_id=by_rank[primary][i]["counter_id"],
                                   reference_sequence_index=i,assignment="matched_to_reference")))
    orphan_count=0
    for row in sorted(records,key=lambda r:(r["received_at_ns"],r["dp_rank"],r["source_line"])):
        if row["source_line"] not in used:
            output.append(([row],dict(method="monotone_anchor_sequence",assignment="unresolved_observation")))
            used.add(row["source_line"]);orphan_count+=1
    assert len(used)==len(records)
    return output,dict(method="monotone_anchor_sequence",version=VERSION,
                        primary_reference_rank=primary,secondary_reference_rank=secondary,
                        positive_records_per_rank={r:len(v) for r,v in by_rank.items()},
                        pairwise=summaries,reference_path_consistency=consistency,
                        sensitivity=sensitivity,rejected_reference_matches=rejected,
                        unresolved_observations=orphan_count,
                        configuration=dict(anchor_candidate_radius_iterations=12,
                                           sensitivity_candidate_radius_iterations=24,
                                           gap_penalty=4.,sensitivity_gap_penalty=3.,
                                           max_segment_cells=1_000_000,
                                           weak_sparse_segment_max_dimension=32,
                                           weak_sparse_segment_min_dimension_exclusive=4),
                        duration_is_alignment_feature=True,
                        warning="Monotone sequence inference, not native global execution IDs. Equal duration after matching is not independent validation.")


def align_records(records,case,full_timed,ordinals):
    expected=case["expected_dp_ranks"]
    if len(expected)==1:
        groups=[([r],dict(method="single_rank")) for r in records]
        report=dict(method="single_rank",version=VERSION)
    elif case["backend"]=="sglang":
        groups,report=sglang_groups(records,expected,full_timed,ordinals)
    else:
        groups,report=vllm_groups(records,expected)
    report["dependencies"]={"numpy":np.__version__,"numba":numba.__version__}
    return groups,report