Added comments

This commit is contained in:
Artur Hyrenko 2026-02-04 20:26:38 +00:00
parent de2336183b
commit 4ef5f31659

View File

@ -15,20 +15,27 @@ from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
# =========================
# Fixed settings (your setup)
# =========================
# Dataset source on Hugging Face Hub and which split to translate.
DATASET_NAME = "PKU-Alignment/PKU-SafeRLHF-30K"
SPLIT = "train"
# Local path to the NLLB model (loaded from disk, not from the Hub).
NLLB_PATH = "/home/hyrenko/Diploma/models/nllb-200-1.3B"
# Output (translated dataset)
# Output directory where the fully translated dataset will be saved.
OUT_FINAL_DIR = "/home/hyrenko/Diploma/datasets/PKU-SafeRLHF-30K_slk_Latn_SFT_DPO_ONLY"
# Language codes for NLLB (source and target).
SRC_LANG = "eng_Latn"
TGT_LANG = "slk_Latn"
# Speed defaults for 2x RTX Titan 24GB
# Performance-oriented defaults for a 2x RTX Titan 24GB setup.
# Notes:
# - MODEL_BATCH_SIZE controls GPU-side translation batching.
# - MAP_BATCH_SIZE controls CPU/RAM side batching when mapping over a dataset shard.
# - LONG_THRESHOLD_CHARS splits very long texts to avoid slow/unstable generation.
# - NUM_BEAMS must be 1 for speed (beam search is expensive).
# - MAX_NEW_TOKENS caps translation length; reducing it can speed up generation.
MODEL_BATCH_SIZE = 32 # if OOM: 24 -> 16
MAP_BATCH_SIZE = 128 # CPU/RAM side
LONG_THRESHOLD_CHARS = 3500 # higher => fewer slow fallbacks
@ -39,12 +46,39 @@ MAX_NEW_TOKENS = 128 # can lower to 96 for speed
# =========================
# Utilities
# =========================
# -----------------------------------------------------------------------------
# normalize_text(x)
# Purpose:
# Converts arbitrary dataset values into a clean string:
# - Handles None values safely.
# - Removes the Unicode replacement char (<28>, U+FFFD) which can appear in noisy text.
# - Strips whitespace.
#
# Why it exists:
# Dataset fields may contain None / non-string types / corrupted characters.
# Translation should always receive a valid string input.
# -----------------------------------------------------------------------------
def normalize_text(x: Any) -> str:
if x is None:
return ""
return str(x).replace("\uFFFD", "").strip()
# -----------------------------------------------------------------------------
# split_text_safely(text, max_chars)
# Purpose:
# Splits long text into smaller chunks to avoid slow generation / memory issues.
#
# Strategy:
# - Normalizes input text.
# - Splits by paragraph boundaries (2+ newlines).
# - Ensures each chunk is <= max_chars.
# - If a paragraph itself is too long, it is cut into fixed-size slices.
#
# Output:
# A list of non-empty chunks; returns [""] for empty input as a safe placeholder.
# -----------------------------------------------------------------------------
def split_text_safely(text: str, max_chars: int) -> List[str]:
text = normalize_text(text)
if not text:
@ -52,6 +86,11 @@ def split_text_safely(text: str, max_chars: int) -> List[str]:
chunks: List[str] = []
# push(piece)
# Internal helper that:
# - ignores empty pieces
# - keeps pieces <= max_chars
# - slices very long pieces into fixed-size parts
def push(piece: str):
piece = piece.strip()
if not piece:
@ -73,6 +112,21 @@ def split_text_safely(text: str, max_chars: int) -> List[str]:
return chunks if chunks else [""]
# -----------------------------------------------------------------------------
# detect_needed_cols_for_sft_dpo(ds)
# Purpose:
# Auto-detect which columns should be translated so the output dataset is
# immediately usable for SFT and/or DPO preparation.
#
# Priority:
# 1) Classic preference schema: prompt + chosen + rejected
# 2) Pairwise schema: prompt + response_0 + response_1
# 3) Fallback: prompt + any response-like columns matching a regex pattern
# 4) Last resort: prompt only
#
# Result:
# A list of columns to translate, minimizing work and storage.
# -----------------------------------------------------------------------------
def detect_needed_cols_for_sft_dpo(ds: Dataset) -> List[str]:
"""
Auto-detect columns needed for SFT->DPO training.
@ -109,6 +163,21 @@ def detect_needed_cols_for_sft_dpo(ds: Dataset) -> List[str]:
raise RuntimeError("Could not detect a 'prompt' column; cannot proceed.")
# -----------------------------------------------------------------------------
# nllb_translate_batch(tokenizer, model, texts)
# Purpose:
# Performs batched translation for a list of short texts via NLLB.
#
# Details:
# - Sets tokenizer.src_lang each time to ensure correct source language.
# - Uses forced_bos_token_id to force decoding in the target language.
# - Tokenizes to max_length=1024 (encoder-side), pads and truncates as needed.
# - Uses model.generate() with max_new_tokens, num_beams, caching for speed.
# - Returns stripped decoded strings.
#
# Important:
# - @torch.inference_mode disables gradients for speed and lower memory.
# -----------------------------------------------------------------------------
@torch.inference_mode()
def nllb_translate_batch(tokenizer, model, texts: List[str]) -> List[str]:
tokenizer.src_lang = SRC_LANG
@ -135,6 +204,17 @@ def nllb_translate_batch(tokenizer, model, texts: List[str]) -> List[str]:
return [o.strip() for o in out]
# -----------------------------------------------------------------------------
# translate_long_text(tokenizer, model, text, sub_batch)
# Purpose:
# Safe fallback path for very long texts that may be slow or unstable if
# translated in one shot.
#
# Strategy:
# - Split the text into <= LONG_THRESHOLD_CHARS chunks (paragraph-aware).
# - Translate chunks in sub-batches to control VRAM use.
# - Join translated chunks with blank lines to preserve paragraph separation.
# -----------------------------------------------------------------------------
def translate_long_text(tokenizer, model, text: str, sub_batch: int) -> str:
text = normalize_text(text)
if not text:
@ -146,6 +226,11 @@ def translate_long_text(tokenizer, model, text: str, sub_batch: int) -> str:
return "\n\n".join([c for c in out_chunks if c])
# -----------------------------------------------------------------------------
# _safe_mkdir(p)
# Purpose:
# Convenience helper to create directories idempotently.
# -----------------------------------------------------------------------------
def _safe_mkdir(p: str):
os.makedirs(p, exist_ok=True)
@ -153,6 +238,23 @@ def _safe_mkdir(p: str):
# =========================
# Worker (one GPU)
# =========================
# -----------------------------------------------------------------------------
# worker_translate(...)
# Purpose:
# Runs translation on a shard of the dataset using a specific GPU.
#
# Key ideas:
# - Uses CUDA_VISIBLE_DEVICES to bind this worker process to a single GPU.
# - Loads the full dataset, then selects only rows belonging to this shard via:
# index % total_shards == shard_id
# This guarantees a complete partition with no missing rows across workers.
# - Translates only the required columns (cols_to_translate).
# - Uses a two-path translation:
# * fast path for short texts (batched GPU translation)
# * slow fallback for very long texts (split + smaller sub-batches)
# - Saves this shards translated dataset to tmp_dir/shard_XX for later merging.
# -----------------------------------------------------------------------------
def worker_translate(
shard_id: int,
gpu_id: int,
@ -161,8 +263,10 @@ def worker_translate(
cols_to_translate: List[str],
total_len: int,
):
# Bind this process to a specific GPU.
os.environ["CUDA_VISIBLE_DEVICES"] = str(gpu_id)
# Global inference optimizations (no gradients).
torch.set_grad_enabled(False)
torch.backends.cudnn.benchmark = True
@ -172,16 +276,18 @@ def worker_translate(
print(f"\n[WORKER {shard_id}] GPU={gpu_id} device={device} dtype={dtype}")
print(f"[WORKER {shard_id}] Translating columns: {cols_to_translate}")
# Load the dataset split (each worker loads it independently).
ds = load_dataset(DATASET_NAME, split=SPLIT)
n = len(ds)
if n != total_len:
print(f"[WORKER {shard_id}] WARN: dataset length mismatch: got {n}, expected {total_len}")
# strict sharding by index => no missing rows
# Strict sharding by original index to avoid missing rows and keep deterministic coverage.
indices = [i for i in range(n) if (i % total_shards) == shard_id]
ds_shard = ds.select(indices)
shard_len = len(ds_shard)
# Load tokenizer + model from local path (NLLB).
tokenizer = AutoTokenizer.from_pretrained(NLLB_PATH, use_fast=True)
model = AutoModelForSeq2SeqLM.from_pretrained(
NLLB_PATH,
@ -190,8 +296,20 @@ def worker_translate(
).to(device)
model.eval()
# Number of CPU-side map batches (used only for progress bar granularity).
total_map_batches = math.ceil(shard_len / MAP_BATCH_SIZE)
# map_fn(batch)
# Purpose:
# Translates the requested columns for a single CPU-side batch.
#
# Workflow per column:
# 1) Normalize all values to strings.
# 2) Split into short_texts and long_texts by LONG_THRESHOLD_CHARS.
# 3) Translate short texts with fast batched translation:
# - internal sub-batches of MODEL_BATCH_SIZE
# 4) Translate long texts individually through translate_long_text().
# 5) Return a dict: {col_name: translated_list} for each requested column.
def map_fn(batch: Dict[str, List[Any]]) -> Dict[str, List[str]]:
out: Dict[str, List[str]] = {}
@ -203,6 +321,7 @@ def worker_translate(
long_texts: List[str] = []
long_pos: List[int] = []
# Partition examples by length so most data goes through the fast path.
for j, t in enumerate(vals):
if (not t) or (len(t) <= LONG_THRESHOLD_CHARS):
short_pos.append(j)
@ -213,7 +332,7 @@ def worker_translate(
translated = [""] * len(vals)
# fast batched translate
# Fast batched translate for short texts (GPU-optimized).
if short_texts:
tr_short: List[str] = []
for k in range(0, len(short_texts), MODEL_BATCH_SIZE):
@ -221,7 +340,7 @@ def worker_translate(
for pos, tr in zip(short_pos, tr_short):
translated[pos] = tr
# slow fallback for very long texts
# Slow fallback for very long texts (split + smaller batches).
for pos, t in zip(long_pos, long_texts):
translated[pos] = translate_long_text(tokenizer, model, t, sub_batch=max(1, MODEL_BATCH_SIZE // 2))
@ -229,7 +348,8 @@ def worker_translate(
return out
# manual loop to show stable progress bar
# Manual loop (instead of a single ds_shard.map) to show a stable progress bar.
# Each loop translates a slice [start:end] of the shard.
parts: List[Dataset] = []
with tqdm(total=total_map_batches, desc=f"GPU{gpu_id} shard{shard_id}", ncols=100) as pbar:
for start in range(0, shard_len, MAP_BATCH_SIZE):
@ -243,8 +363,10 @@ def worker_translate(
parts.append(part)
pbar.update(1)
# Concatenate translated parts into a single shard dataset.
ds_tr = concatenate_datasets(parts)
# Save this shard to disk for later merge.
shard_path = os.path.join(tmp_dir, f"shard_{shard_id:02d}")
_safe_mkdir(shard_path)
ds_tr.save_to_disk(shard_path)
@ -254,21 +376,44 @@ def worker_translate(
# =========================
# Main (one-click)
# =========================
# -----------------------------------------------------------------------------
# main()
# Purpose:
# One-click launcher for 2-GPU translation + shard merge.
#
# Steps:
# 1) Parse CLI args:
# --resume: skip workers for shards that already exist and only merge.
# 2) Validate:
# - NLLB model path exists
# - at least 2 GPUs are available (script expects exactly 2 workers here)
# 3) Load dataset metadata (length + schema) and determine which columns to translate.
# 4) Spawn 2 worker processes (GPU 0 and GPU 1) using spawn start method.
# 5) After workers finish, merge shard datasets and restore original row order.
# 6) Save the final merged dataset to OUT_FINAL_DIR and verify row count.
# -----------------------------------------------------------------------------
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--resume", action="store_true", help="Skip existing shards and only merge.")
args = ap.parse_args()
# Ensure the local NLLB path is valid before starting.
if not os.path.isdir(NLLB_PATH):
raise SystemExit(f"[ERROR] NLLB model path not found: {NLLB_PATH}")
# This script is designed for a 2-GPU setup (two workers).
if torch.cuda.device_count() < 2:
raise SystemExit(f"[ERROR] Need 2 GPUs; found: {torch.cuda.device_count()}")
# Prepare output directories.
_safe_mkdir(OUT_FINAL_DIR)
tmp_dir = os.path.join(OUT_FINAL_DIR, "_tmp_shards")
_safe_mkdir(tmp_dir)
# Load dataset once in the main process to:
# - get total length n
# - auto-detect which columns are necessary for SFT/DPO
print("[INFO] Loading dataset metadata...")
ds = load_dataset(DATASET_NAME, split=SPLIT)
n = len(ds)
@ -281,12 +426,15 @@ def main():
print(f"[INFO] Params: MODEL_BATCH_SIZE={MODEL_BATCH_SIZE}, MAP_BATCH_SIZE={MAP_BATCH_SIZE}, "
f"LONG_THRESHOLD_CHARS={LONG_THRESHOLD_CHARS}, NUM_BEAMS={NUM_BEAMS}, MAX_NEW_TOKENS={MAX_NEW_TOKENS}")
# Two workers => two shards, pinned to GPU 0 and GPU 1.
gpus = [0, 1]
total_shards = 2
# Use "spawn" for CUDA safety in multiprocessing.
mp.set_start_method("spawn", force=True)
procs = []
# Start workers unless resume mode skips an existing shard.
for shard_id, gpu_id in enumerate(gpus):
shard_path = os.path.join(tmp_dir, f"shard_{shard_id:02d}")
if args.resume and os.path.isdir(shard_path):
@ -300,12 +448,16 @@ def main():
p.start()
procs.append(p)
# Wait for workers and fail fast if any worker exits with a non-zero code.
for p in procs:
p.join()
if p.exitcode != 0:
raise SystemExit("[ERROR] A worker failed. See logs above.")
# merge
# Merge step:
# - load each shard dataset from disk
# - concatenate them
# - restore original ordering (because sharding interleaves indices)
print("\n[INFO] Merging shards...")
shard_dirs = [os.path.join(tmp_dir, f"shard_{i:02d}") for i in range(total_shards)]
for sd in shard_dirs:
@ -315,17 +467,21 @@ def main():
shards = [Dataset.load_from_disk(sd) for sd in shard_dirs]
merged = concatenate_datasets(shards)
# restore original order
# Restore original order:
# Each shard contains indices satisfying (i % total_shards == shard_id).
merged_indices = []
for shard_id in range(total_shards):
merged_indices.extend([i for i in range(n) if (i % total_shards) == shard_id])
# Add an ordering column, sort by it, then remove it.
merged = merged.add_column("orig_index", merged_indices)
merged = merged.sort("orig_index").remove_columns(["orig_index"])
# Safety check: ensure no missing rows after merge.
if len(merged) != n:
raise SystemExit(f"[ERROR] Length mismatch after merge: merged={len(merged)} expected={n}")
# Save the final translated dataset.
print("[INFO] Saving final dataset...")
merged.save_to_disk(OUT_FINAL_DIR)