generate-embeddings.py

# /// script
# requires-python = ">=3.10"
# dependencies = [
#     "datasets",
#     "sentence-transformers>=5.0.0",
#     "torch",
#     "numpy",
#     "pillow",
#     "einops",
#     "huggingface-hub>=1.12",
# ]
# ///
"""
Generate embeddings for a Hugging Face dataset (text OR images) with sentence-transformers,
and push the result back to the Hub as a new dataset with an `embeddings` column.

This is the simple, ergonomic default. It runs as one command on the bare uv image, on CPU
or any GPU flavor. For maximum throughput on large *decoder* embedding models (e.g.
Qwen3-Embedding), see the vLLM variant; to get a searchable vector index as a Hub dataset,
see the Lance variant.

PROMPTS (retrieval correctness — read this):
    Many embedding models need a DIFFERENT prefix/instruction for documents vs queries, and
    getting it wrong silently degrades retrieval. This script embeds a *document corpus* by
    default, via sentence-transformers' native encode_document()/encode_query() (which also
    route Router models by task), picking the right document convention for you:
      1. the model's REGISTERED prompt if it ships one (e.g. Qwen3-Embedding) — selected
         natively by encode_document/encode_query, else
      2. a small built-in table of well-known families (e5, nomic, bge), else
      3. no prefix.
    Heads-up: current sentence-transformers injects a placeholder prompts dict
    {"query": "", "document": ""} even for models that register NOTHING — so e5 ("passage: "),
    nomic ("search_document: ") etc. look prompt-less via `.prompts`; their real prefixes live
    only in the model card. The built-in table handles that. Override with --prompt '<prefix>'
    or --prompt-name <registered-name>; embed a query set with --query-mode; force no prefix
    with --prompt ''. The chosen prompt is logged and recorded in the dataset card.

Benchmarks (20k rows, seq-cap 512): all-MiniLM-L6-v2 ~900 rows/s on an L4 (~$0.24/1M rows);
bge-base-en-v1.5 ~120 rows/s. L4 is the cheapest flavor for these encoder models.

Examples:
    # Text (default). Document convention auto-picked.
    hf jobs uv run --flavor l4x1 -s HF_TOKEN generate-embeddings.py \\
        stanfordnlp/imdb  your-name/imdb-embeddings \\
        --column text --model sentence-transformers/all-MiniLM-L6-v2

    # e5: docs auto-get "passage: ". (--prompt 'passage: ' would be the explicit form.)
    hf jobs uv run --flavor l4x1 -s HF_TOKEN generate-embeddings.py \\
        stanfordnlp/imdb  your-name/imdb-e5 --model intfloat/multilingual-e5-large

    # Images (CLIP) — prompts don't apply.
    hf jobs uv run --flavor l4x1 -s HF_TOKEN generate-embeddings.py \\
        your-name/photos  your-name/photos-embeddings \\
        --modality image --column image --model clip-ViT-B-32

    # Test on a small slice first, keep the output private
    hf jobs uv run --flavor l4x1 -s HF_TOKEN generate-embeddings.py \\
        stanfordnlp/imdb  your-name/imdb-emb --max-samples 100 --private

FAN-OUT (multi-job) MODE:
    Split one embedding run across N Jobs with --num-shards/--shard-index (or the RANK /
    NUM_SHARDS / RUN_ID / OUTPUT_BUCKET env vars, so a launcher only varies RANK per job —
    see launch-embedding-fleet.py). Each shard takes an exact, non-overlapping contiguous
    slice, writes its rows to the run's bucket as runs/<run-id>/data/<rank>.parquet (object
    PUTs — no repo-commit contention), and heartbeats progress to
    runs/<run-id>/status/<rank>.json. consolidate-shards.py merges shards into the final
    dataset with one commit. Re-running a failed rank overwrites only its own files (resume).
    Note: --max-samples applies BEFORE sharding, so a capped run still partitions exactly.
    Each rank still downloads the full split before slicing — fine at few-million-row scale;
    for larger corpora shard at the file level or stream instead.
"""
import argparse
import logging
import os
import re
import sys
import time

logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
logger = logging.getLogger("generate-embeddings")


def find_batch_size(model, sample, normalize, candidates=(32, 64, 128, 256)):
    """Probe for the fastest batch that fits (used by --batch-size auto). Throughput is NOT
    monotonic in batch size, so we time a few on a warmup sample and keep the fastest that doesn't
    OOM. Why bigger isn't better: for text, larger batches pad to the longest member + add overhead;
    for images, the ViT forward already saturates the GPU by ~batch 32. Works for text and images."""
    import time
    import torch
    warm = sample[: min(1024, len(sample))]
    try:  # one untimed warmup so cudnn autotune doesn't penalise the first probe
        model.encode(warm[:32], batch_size=32, show_progress_bar=False,
                     convert_to_numpy=True, normalize_embeddings=normalize)
    except Exception:
        pass
    best_bs, best_rps = candidates[0], 0.0
    for bs in candidates:
        try:
            if torch.cuda.is_available():
                torch.cuda.empty_cache()
                torch.cuda.synchronize()
            t = time.perf_counter()
            model.encode(warm, batch_size=bs, show_progress_bar=False,
                         convert_to_numpy=True, normalize_embeddings=normalize)
            if torch.cuda.is_available():
                torch.cuda.synchronize()
            rps = len(warm) / (time.perf_counter() - t)
            logger.info(f"  auto-batch probe bs={bs}: {rps:.0f} rows/s")
            if rps > best_rps:
                best_rps, best_bs = rps, bs
        except RuntimeError as e:
            if "out of memory" in str(e).lower():
                logger.info(f"  auto-batch bs={bs} OOM → stopping probe")
                if torch.cuda.is_available():
                    torch.cuda.empty_cache()
                break
            raise
    logger.info(f"auto-batch chose bs={best_bs} ({best_rps:.0f} rows/s on warmup)")
    return best_bs


def known_convention(model_id):
    """Best-effort (query_prefix, doc_prefix) for common families whose convention is
    documented in the model card but NOT registered in config_sentence_transformers.json.
    Returns None if unknown. Overridable with --prompt / --no-auto-prompt.

    Verified 2026-07-03 on HF Jobs: of e5 / nomic / bge-en / bge-m3 / Qwen3-Embedding, only
    Qwen3-Embedding registers real ST prompts; the rest ship none and rely on manual prefixes.
    """
    m = model_id.lower()
    # Instruction-style embedders (e5-*-instruct, gte-Qwen, ...): prefer the model's REGISTERED
    # prompt or an explicit --prompt; don't guess a literal prefix.
    if "instruct" in m:
        return None
    if "nomic-embed-text" in m:
        return ("search_query: ", "search_document: ")
    if "bge-m3" in m:  # bge-m3 uses no prompts
        return ("", "")
    # e5 family (e5-base/large/small, multilingual-e5-*), boundaried so e.g. "table5" or a
    # model with "e5" mid-word can't silently pick up "query:/passage:" prefixes.
    if re.search(r"(^|[/_-])e5([_-]|$)", m):
        return ("query: ", "passage: ")
    if "bge" in m and "-en" in m:  # English bge retrieval: query instruction, docs raw
        return ("Represent this sentence for searching relevant passages: ", "")
    return None


def resolve_prompt(model, model_id, is_query, args):
    """Decide the EXPLICIT prefix to pass to encode_query()/encode_document(), or None to let the
    native method choose. sentence-transformers' encode_query/encode_document already select the
    model's REGISTERED query/document prompt and set the Router task — we lean on that, and only
    supply a prefix ourselves for (a) explicit --prompt/--prompt-name, (b) the known-family table
    covering models that register nothing (e5, nomic, bge-en — their prefixes live only in the
    model card, so the native fallback would silently apply NO prefix)."""
    registered = dict(getattr(model, "prompts", {}) or {})
    # Current sentence-transformers injects a placeholder {"query":"","document":""} for models
    # with no config prompts; only non-empty values are real conventions.
    real = {k: v for k, v in registered.items() if v}
    logger.info(f"Registered prompts: {registered} · real (non-empty): {real or 'none'} · "
                f"default_prompt_name={getattr(model, 'default_prompt_name', None)}")
    side = "query" if is_query else "document"

    if args.prompt is not None:  # includes --prompt '' to force no prefix
        logger.info(f"Prompt: raw --prompt → {args.prompt!r}")
        return args.prompt
    if args.prompt_name:
        if args.prompt_name not in registered:
            logger.error(f"--prompt-name {args.prompt_name!r} not registered ({list(registered)}); "
                         f"use --prompt '<raw prefix>' instead.")
            sys.exit(1)
        logger.info(f"Prompt: registered prompt_name={args.prompt_name!r} → {registered[args.prompt_name]!r}")
        return registered[args.prompt_name]
    native_keys = ("query",) if is_query else ("document", "passage", "corpus")
    if any(real.get(k) for k in native_keys):
        # Model ships a real prompt for this side (e.g. Qwen3 query) → encode_query/encode_document
        # selects it natively (and routes Router models by task).
        logger.info(f"Prompt: model-registered — selected natively by encode_{side}()")
        return None

    kc = known_convention(model_id)
    if kc is not None:
        chosen = kc[0] if is_query else kc[1]
        if args.no_auto_prompt:
            if chosen:
                logger.warning(f"--no-auto-prompt set: NOT applying the known {side} prefix {chosen!r} for "
                               f"{model_id}. Retrieval may degrade unless you pass --prompt.")
            return ""
        logger.info(f"Prompt: known-family {side} prefix → {chosen!r} (override with --prompt)"
                    if chosen else f"Prompt: known-family → no {side} prefix needed")
        return chosen

    logger.info(f"Prompt: none registered or known for {model_id} — encode_{side}() applies no prefix. "
                f"If it's a retrieval model needing a query/document prefix, pass --prompt.")
    return None


def sniff_token_lengths(model, texts, max_seq_len, sample=512):
    """Tokenize a sample to report the token-length distribution + how much --max-seq-len truncates,
    and return the median length (used to pick the auto-batch candidate range: short texts under-use
    the GPU at small batch, long texts waste compute on padding). Text only; returns None on failure."""
    try:
        tok = model.tokenizer
    except Exception:
        return None
    s = texts[: min(sample, len(texts))]
    lens = sorted(len(tok.encode(t, add_special_tokens=True)) for t in s)
    n = len(lens)
    if not n:
        return None
    median, p90, mx = lens[n // 2], lens[min(n - 1, int(n * 0.9))], lens[-1]
    pct_over = 100 * sum(1 for L in lens if L > max_seq_len) / n
    note = (f" → {pct_over:.0f}% exceed --max-seq-len {max_seq_len} and are truncated "
            f"(raise it to keep more, at higher cost/slower)" if pct_over >= 5
            else f" (all within --max-seq-len {max_seq_len})")
    logger.info(f"Token lengths (sample {n}): median {median}, p90 {p90}, max {mx}{note}")
    return median


def put_bucket_files(bucket_id, add, max_retries=3):
    """batch_bucket_files with a short retry — bucket PUTs are cheap object writes but can
    flake transiently; shard data must not be lost to a blip."""
    from huggingface_hub import batch_bucket_files
    for attempt in range(1, max_retries + 1):
        try:
            batch_bucket_files(bucket_id, add=add)
            return True
        except Exception as e:
            if attempt == max_retries:
                raise
            logger.warning(f"bucket PUT attempt {attempt}/{max_retries} failed: {e}; retrying in {5 * attempt}s")
            time.sleep(5 * attempt)


class StatusReporter:
    """Fan-out worker heartbeat: PUTs runs/<run-id>/status/<rank>.json to the run bucket,
    throttled to one write per ~30s (last-writer-wins per key, so N workers never contend).
    A failed status write must never kill a paid embedding run — errors are logged, not raised."""

    def __init__(self, bucket, run_id, rank, rows_total, tokens_per_row=None):
        self.bucket, self.run_id, self.rank = bucket, run_id, rank
        self.rows_total, self.tokens_per_row = rows_total, tokens_per_row
        self.started_at = time.time()
        self.rows_done = 0
        self._last_write = 0.0

    def report(self, rows_done=None, state="running", force=False):
        import json
        if rows_done is not None:
            self.rows_done = rows_done
        now = time.time()
        if not force and now - self._last_write < 30:
            return
        elapsed = max(now - self.started_at, 1e-6)
        payload = {
            "run_id": self.run_id,
            "rank": self.rank,
            "state": state,
            "rows_done": self.rows_done,
            "rows_total": self.rows_total,
            "tokens_done_est": int(self.rows_done * self.tokens_per_row) if self.tokens_per_row else None,
            "rows_per_sec": round(self.rows_done / elapsed, 1),
            "started_at": self.started_at,
            "updated_at": now,
            "job_id": os.environ.get("JOB_ID"),
        }
        dest = f"runs/{self.run_id}/status/{self.rank:05d}.json"
        try:
            put_bucket_files(self.bucket, [(json.dumps(payload).encode(), dest)], max_retries=2)
            self._last_write = now
        except Exception as e:
            logger.warning(f"status write skipped ({e})")


def run_streaming_shard(ds, model, prompt_str, args):
    """Streaming fan-out worker: iterate this rank's FILE-shard, encode in chunks, and flush
    parquet PARTS to the bucket every ~250k rows, so memory and disk stay bounded no matter
    how big the shard is. Output keys: runs/<run-id>/data/<rank>.part<p>.parquet — the
    consolidator accepts both this and row mode's single <rank>.parquet naming."""
    import itertools
    import pyarrow as pa
    import pyarrow.parquet as pq

    encode_fn = model.encode_query if args.query_mode else model.encode_document
    encode_kwargs = {"prompt": prompt_str} if prompt_str is not None else {}

    def clean(t):
        return t if isinstance(t, str) and t.strip() else " "

    # Buffer a head sample for token sniffing + the auto-batch probe, then chain it back.
    it = iter(ds)
    head = list(itertools.islice(it, 1024))
    if not head:
        logger.error(f"File-shard {args.shard_index} yielded no rows.")
        sys.exit(1)
    if args.column not in head[0]:
        logger.error(f"Column {args.column!r} not in rows. Available: {sorted(head[0])}")
        sys.exit(1)
    if args.output_column in head[0]:
        logger.error(f"Output column {args.output_column!r} already exists — choose another --output-column.")
        sys.exit(1)
    head_texts = [clean(r[args.column]) for r in head]
    median_tok = sniff_token_lengths(model, head_texts, args.max_seq_len)
    if str(args.batch_size).lower() == "auto":
        if median_tok is None or median_tok >= 256:
            candidates = (32, 64, 128, 256)
        elif median_tok >= 64:
            candidates = (64, 128, 256, 512)
        else:
            candidates = (128, 256, 512, 1024)
        batch_size = find_batch_size(model, head_texts, args.normalize, candidates=candidates)
    else:
        batch_size = int(args.batch_size)

    reporter = StatusReporter(args.output_bucket, args.run_id, args.shard_index,
                              rows_total=None, tokens_per_row=median_tok)
    reporter.report(0, force=True)

    chunk_rows, part_rows = 25_000, 250_000
    prog = {"part_idx": 0, "rows_done": 0}
    buf_rows, buf_texts, part_buf = [], [], []

    def flush_part():
        if not part_buf:
            return
        path = f"/tmp/part-{args.shard_index:05d}-{prog['part_idx']:04d}.parquet"
        pq.write_table(pa.Table.from_pylist(part_buf), path,
                       row_group_size=25_000, write_page_index=True)
        dest = f"runs/{args.run_id}/data/{args.shard_index:05d}.part{prog['part_idx']:04d}.parquet"
        logger.info(f"Uploading {len(part_buf):,}-row part → {args.output_bucket}/{dest}")
        put_bucket_files(args.output_bucket, [(path, dest)])
        os.remove(path)
        part_buf.clear()
        prog["part_idx"] += 1

    def encode_chunk():
        if not buf_rows:
            return
        emb = encode_fn(buf_texts, batch_size=batch_size, show_progress_bar=False,
                        convert_to_numpy=True, normalize_embeddings=args.normalize, **encode_kwargs)
        for r, e in zip(buf_rows, emb):
            r[args.output_column] = e.tolist()
        part_buf.extend(buf_rows)
        prog["rows_done"] += len(buf_rows)
        buf_rows.clear()
        buf_texts.clear()
        reporter.report(prog["rows_done"])
        if len(part_buf) >= part_rows:
            flush_part()

    t0 = time.perf_counter()
    try:
        for row in itertools.chain(head, it):
            buf_rows.append(dict(row))
            buf_texts.append(clean(row[args.column]))
            if len(buf_rows) >= chunk_rows:
                encode_chunk()
        encode_chunk()
        flush_part()
    except Exception:
        reporter.report(state="error", force=True)
        raise
    secs = time.perf_counter() - t0
    logger.info(f"Embedded {prog['rows_done']:,} rows in {secs:.0f}s "
                f"({prog['rows_done'] / max(secs, 1e-6):.0f} rows/s), {prog['part_idx']} part(s)")
    reporter.report(prog["rows_done"], state="done", force=True)
    logger.info(f"✅ streaming shard {args.shard_index + 1}/{args.num_shards} of run {args.run_id} uploaded")


def main():
    p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
    p.add_argument("input_dataset", help="Input dataset ID on the Hugging Face Hub")
    p.add_argument("output_dataset", help="Output dataset ID to create on the Hub")
    p.add_argument("--model", default="sentence-transformers/all-MiniLM-L6-v2",
                   help="sentence-transformers model (text or CLIP image model)")
    p.add_argument("--modality", choices=["text", "image"], default="text")
    p.add_argument("--column", default="text", help="Input column (text string, or image)")
    p.add_argument("--output-column", default="embeddings")
    p.add_argument("--config", default=None, help="Dataset config name (e.g. wikipedia needs one)")
    p.add_argument("--split", default="train")
    p.add_argument("--max-samples", type=int, default=None, help="Limit rows (for testing)")
    p.add_argument("--batch-size", default="auto",
                   help="'auto' probes for the fastest batch that fits, or pass an int")
    p.add_argument("--prompt", default=None,
                   help="Raw prefix to prepend to every text (e.g. 'passage: '). Highest precedence. "
                        "Use --prompt '' to force NO prefix.")
    p.add_argument("--prompt-name", default=None,
                   help="Name of a prompt REGISTERED by the model (e.g. 'query'); errors if not registered.")
    p.add_argument("--query-mode", action="store_true",
                   help="Embed inputs as QUERIES, not documents (flips the auto-picked convention).")
    p.add_argument("--no-auto-prompt", action="store_true",
                   help="Disable the built-in known-family prefix table (still honours registered prompts).")
    p.add_argument("--max-seq-len", type=int, default=512,
                   help="Truncate text to this many tokens (predictable cost; RAG-typical)")
    p.add_argument("--normalize", action="store_true", default=True)
    p.add_argument("--no-normalize", dest="normalize", action="store_false")
    p.add_argument("--private", action="store_true", help="Make the output dataset private")
    # Fan-out mode (see docstring). Env-defaulted so a launcher drives workers purely via env.
    # Defaults stay raw strings; int conversion happens after parse so a malformed env var
    # reaches p.error() instead of blowing up argparse construction.
    p.add_argument("--num-shards", default=os.environ.get("NUM_SHARDS"),
                   help="Fan-out: total number of shards (env NUM_SHARDS). Enables shard mode.")
    p.add_argument("--shard-index", default=os.environ.get("RANK"),
                   help="Fan-out: this worker's shard index, 0-based (env RANK).")
    p.add_argument("--output-bucket", default=os.environ.get("OUTPUT_BUCKET"),
                   help="Fan-out: bucket for shard parquets + status (env OUTPUT_BUCKET).")
    p.add_argument("--run-id", default=os.environ.get("RUN_ID"),
                   help="Fan-out: run identifier grouping shards under runs/<run-id>/ (env RUN_ID).")
    p.add_argument("--streaming", action="store_true",
                   default=os.environ.get("STREAMING") == "1",
                   help="Fan-out: stream the dataset and shard at the FILE level (env STREAMING=1). "
                        "Each rank downloads only its own files — use for very big datasets. "
                        "Text modality only; incompatible with --max-samples.")
    p.add_argument("--revision", default=os.environ.get("REVISION"),
                   help="Input dataset revision (commit sha). Pin this in fan-out runs so every "
                        "rank slices the identical snapshot (env REVISION).")
    args = p.parse_args()

    def int_or_error(val, name):
        if val is None:
            return None
        try:
            return int(val)
        except (TypeError, ValueError):
            p.error(f"{name} must be an integer (got {val!r}).")

    args.num_shards = int_or_error(args.num_shards, "--num-shards / NUM_SHARDS")
    args.shard_index = int_or_error(args.shard_index, "--shard-index / RANK")

    sharded = args.num_shards is not None
    if sharded:
        if args.num_shards < 1:
            p.error(f"--num-shards must be >= 1 (got {args.num_shards}).")
        if args.shard_index is None or not 0 <= args.shard_index < args.num_shards:
            p.error(f"--shard-index must be in [0, {args.num_shards}) when --num-shards is set "
                    f"(got {args.shard_index}).")
        if not (args.output_bucket and args.run_id):
            p.error("fan-out mode needs --output-bucket and --run-id (env OUTPUT_BUCKET / RUN_ID).")
    elif args.shard_index is not None:
        p.error("--shard-index requires --num-shards.")

    if args.streaming:
        if not sharded:
            p.error("--streaming is a fan-out mode — it needs --num-shards/--shard-index.")
        if args.modality != "text":
            p.error("--streaming currently supports text modality only.")
        if args.max_samples:
            p.error("--max-samples is incompatible with --streaming (use row mode for capped test runs).")

    import torch
    from datasets import load_dataset
    from huggingface_hub import DatasetCard, login
    from sentence_transformers import SentenceTransformer

    token = os.environ.get("HF_TOKEN")
    if token:
        login(token=token)
    if not torch.cuda.is_available():
        logger.warning("No CUDA — running on CPU (much slower). Prefer a GPU flavor, e.g. --flavor l4x1.")

    logger.info(f"Loading {args.input_dataset} [{args.split}]"
                + (f" @ {args.revision[:12]}" if args.revision else "")
                + (" (streaming)" if args.streaming else ""))
    load_kwargs = {"split": args.split, "streaming": args.streaming}
    if args.revision:
        load_kwargs["revision"] = args.revision
    ds = (load_dataset(args.input_dataset, args.config, **load_kwargs) if args.config
          else load_dataset(args.input_dataset, **load_kwargs))
    if ds.column_names is not None:
        if args.column not in ds.column_names:
            logger.error(f"Column {args.column!r} not found. Available: {ds.column_names}")
            sys.exit(1)
        if args.output_column in ds.column_names:
            logger.error(f"Output column {args.output_column!r} already exists — choose another --output-column.")
            sys.exit(1)
    if args.max_samples and not args.streaming:
        ds = ds.select(range(min(args.max_samples, len(ds))))
    if sharded and args.streaming:
        # File-level split: each rank reads ONLY its own data files. Sizes vary per rank and
        # rows_total is unknown upfront; correctness (exact, deterministic, idempotent) holds
        # as long as every rank pins the same --revision.
        if ds.n_shards < args.num_shards:
            logger.error(f"Dataset has {ds.n_shards} file shard(s) < --num-shards {args.num_shards}. "
                         f"Lower --num-shards or use row mode.")
            sys.exit(1)
        ds = ds.shard(num_shards=args.num_shards, index=args.shard_index)
        logger.info(f"Fan-out file-shard {args.shard_index + 1}/{args.num_shards} of run {args.run_id} "
                    f"({ds.n_shards} file(s) for this rank)")
    elif sharded:
        # Contiguous slices keep row order reconstructable at consolidation. Note the whole
        # split was still downloaded above — acceptable at few-M rows, not at corpus scale.
        ds = ds.shard(num_shards=args.num_shards, index=args.shard_index, contiguous=True)
        logger.info(f"Fan-out shard {args.shard_index + 1}/{args.num_shards} of run {args.run_id}")
        if len(ds) == 0:
            logger.error(f"Shard {args.shard_index} is empty — num_shards exceeds the row count. "
                         f"Lower --num-shards (or raise --max-samples).")
            sys.exit(1)
    if not args.streaming:
        logger.info(f"{len(ds)} rows; modality={args.modality}")

    device = "cuda" if torch.cuda.is_available() else "cpu"
    model = SentenceTransformer(args.model, device=device, trust_remote_code=True)
    if args.modality == "text" and getattr(model, "max_seq_length", None):
        model.max_seq_length = min(model.max_seq_length, args.max_seq_len)
    dim = model.get_sentence_embedding_dimension()
    logger.info(f"Model {args.model} on {device}; dim={dim}")

    # Prompt handling — many retrieval models need a query vs document/passage prefix (text only).
    prompt_str = None  # None = let encode_query/encode_document choose natively
    if args.modality == "text":
        prompt_str = resolve_prompt(model, args.model, is_query=args.query_mode, args=args)
        if args.streaming:
            run_streaming_shard(ds, model, prompt_str, args)
            return
        items = [t if isinstance(t, str) and t.strip() else " " for t in ds[args.column]]
    else:
        if args.prompt or args.prompt_name:
            logger.warning("--prompt/--prompt-name ignored for image modality.")
        items = [im.convert("RGB") if hasattr(im, "convert") else im for im in ds[args.column]]

    median_tok = sniff_token_lengths(model, items, args.max_seq_len) if args.modality == "text" else None

    if str(args.batch_size).lower() == "auto":
        # Pick the probe range from the data shape. Images: the ViT forward saturates the GPU by
        # ~batch 32, so bigger only adds memory — probe low. Text: short texts under-use the GPU at
        # small batch (probe bigger); long texts pad-waste at big batch (stay modest). Probe verifies.
        if args.modality == "image":
            candidates = (32, 64, 128)
        elif median_tok is None or median_tok >= 256:
            candidates = (32, 64, 128, 256)
        elif median_tok >= 64:
            candidates = (64, 128, 256, 512)
        else:
            candidates = (128, 256, 512, 1024)
        logger.info(f"Finding batch size (--batch-size auto; candidates {candidates})...")
        batch_size = find_batch_size(model, items, args.normalize, candidates=candidates)
    else:
        batch_size = int(args.batch_size)

    # Text goes through encode_query/encode_document (native registered-prompt selection + Router
    # task routing); our resolved prefix, when not None, overrides via prompt=. Images use encode().
    if args.modality == "text":
        encode_fn = model.encode_query if args.query_mode else model.encode_document
        encode_kwargs = {"prompt": prompt_str} if prompt_str is not None else {}
    else:
        encode_fn = model.encode
        encode_kwargs = {}
    reporter = None
    if sharded:
        reporter = StatusReporter(args.output_bucket, args.run_id, args.shard_index,
                                  rows_total=len(items), tokens_per_row=median_tok)
        reporter.report(0, force=True)

    t0 = time.perf_counter()
    try:
        if reporter is None:
            emb = encode_fn(items, batch_size=batch_size, show_progress_bar=True,
                            convert_to_numpy=True, normalize_embeddings=args.normalize, **encode_kwargs)
        else:
            # Macro-chunks so the heartbeat can report rows_done between encode calls.
            # Chunking doesn't change results — batching happens at batch_size regardless.
            import numpy as np
            chunk_rows = 25_000
            parts = []
            for start in range(0, len(items), chunk_rows):
                chunk = items[start:start + chunk_rows]
                parts.append(encode_fn(chunk, batch_size=batch_size, show_progress_bar=True,
                                       convert_to_numpy=True, normalize_embeddings=args.normalize,
                                       **encode_kwargs))
                reporter.report(start + len(chunk))
            emb = np.concatenate(parts) if len(parts) > 1 else parts[0]
    except Exception:
        if reporter:
            reporter.report(state="error", force=True)
        raise
    secs = time.perf_counter() - t0
    logger.info(f"Embedded {len(items)} in {secs:.1f}s ({len(items)/secs:.0f} rows/s), dim={dim}")

    if sharded:
        # Shard mode: no repo commit here (N workers committing → 412 contention). Write this
        # rank's parquet to the run bucket; consolidate-shards.py makes the single final commit.
        #
        # Build the embedding column as zero-copy float32 Arrow — NEVER as Python floats.
        # `[e.tolist() for e in emb]` on an 800k-row shard is ~10 GB of PyFloat objects and
        # swap-thrashed L4 workers into multi-hour silent stalls (wiki fleet, 2026-07-09).
        import numpy as np
        import pyarrow as pa
        import pyarrow.parquet as pq
        reporter.report(len(items), state="writing", force=True)
        try:
            emb32 = np.ascontiguousarray(emb, dtype=np.float32)
            emb_col = pa.FixedSizeListArray.from_arrays(pa.array(emb32.ravel()), emb32.shape[1])
            table = ds.with_format("arrow")[:].append_column(args.output_column, emb_col)
            out_path = f"/tmp/shard-{args.shard_index:05d}.parquet"
            dest = f"runs/{args.run_id}/data/{args.shard_index:05d}.parquet"
            # Small row groups + page index or the dataset viewer can't random-access:
            # pyarrow's default ~1M-row groups ≈ multi-GB with an embeddings column
            # ("Scan size limit exceeded"). ~25k rows ≈ ~100MB here.
            pq.write_table(table, out_path, row_group_size=25_000, write_page_index=True)
            logger.info(f"Uploading shard parquet → {args.output_bucket}/{dest}")
            put_bucket_files(args.output_bucket, [(out_path, dest)])
        except Exception:
            reporter.report(state="error", force=True)
            raise
        reporter.report(len(items), state="done", force=True)
        logger.info(f"✅ shard {args.shard_index + 1}/{args.num_shards} of run {args.run_id} uploaded")
        return

    ds = ds.add_column(args.output_column, [e.tolist() for e in emb])

    # For the card: record the effective prefix (explicit, else the model's registered one).
    side_keys = ("query",) if args.query_mode else ("document", "passage", "corpus")
    effective = prompt_str if prompt_str is not None else next(
        (v for k in side_keys if (v := (getattr(model, "prompts", {}) or {}).get(k))), "")
    prompt_line = f"`{effective}`" if effective else "(none)"
    # Canonical provenance stamp (see AGENTS.md): Jobs claim gated on JOB_ID, set by HF Jobs in-container.
    script_url = "https://huggingface.co/datasets/uv-scripts/embeddings/raw/main/generate-embeddings.py"
    on_jobs = os.environ.get("JOB_ID") is not None
    hw = os.environ.get("ACCELERATOR") or ""
    origin = (
        "Produced on [Hugging Face Jobs](https://huggingface.co/docs/huggingface_hub/guides/jobs)"
        + (f" (`{hw}`)" if hw else "")
    ) if on_jobs else "Generated"
    jobs_tag = "\n- hf-jobs" if on_jobs else ""
    card = DatasetCard(
        f"---\ntags:\n- embeddings\n- uv-script\n- generated{jobs_tag}\n---\n\n"
        f"# {args.output_dataset}\n\n"
        f"Embeddings of [`{args.input_dataset}`](https://huggingface.co/datasets/{args.input_dataset}) "
        f"column `{args.column}`.\n\n"
        f"- Model: [`{args.model}`](https://huggingface.co/{args.model}) (dim {dim})\n"
        f"- Column: `{args.output_column}`  ·  normalized: {args.normalize}\n"
        f"- Prompt prepended ({'query' if args.query_mode else 'document'} side): {prompt_line}\n\n"
        f"## Reproduction\n\n"
        f"{origin} with the [`generate-embeddings.py`]({script_url}) recipe "
        f"from [uv-scripts](https://huggingface.co/uv-scripts). Run it yourself:\n\n"
        f"```bash\nhf jobs uv run {script_url} \\\n"
        f"    {args.input_dataset} <output-dataset> --column {args.column} --model {args.model}\n```\n"
    )
    # Retry the push with an XET-disable fallback: a transient upload failure here would
    # otherwise lose the whole (paid) embedding run.
    logger.info(f"Pushing to {args.output_dataset} (private={args.private})")
    max_retries = 3
    for attempt in range(1, max_retries + 1):
        try:
            if attempt > 1:
                logger.warning("Disabling XET (fallback to HTTP upload)")
                os.environ["HF_HUB_DISABLE_XET"] = "1"
            ds.push_to_hub(args.output_dataset, private=args.private)
            break
        except Exception as e:
            logger.error(f"Upload attempt {attempt}/{max_retries} failed: {e}")
            if attempt < max_retries:
                delay = 30 * (2 ** (attempt - 1))
                logger.info(f"Retrying in {delay}s...")
                time.sleep(delay)
            else:
                logger.error("All upload attempts failed. Results are lost.")
                sys.exit(1)
    try:
        card.push_to_hub(args.output_dataset, repo_type="dataset")
    except Exception as e:
        logger.warning(f"card push skipped: {e}")
    logger.info(f"✅ https://huggingface.co/datasets/{args.output_dataset}")


if __name__ == "__main__":
    main()