scripts/test_ddp_vlm_variable_shape.py

#!/usr/bin/env python3
"""Stress native DDP with deliberately unequal real ChartQA image lengths.

The prior formal OPD run failed in DeepSpeed ZeRO-2's gradient-partition
collective after one rank processed a 17-patch ChartQA image while the other
ranks had much shorter inputs.  This bounded regression test gives rank 0 the
two largest hard-correct images and every other rank two of the smallest ones.
It executes the same eight-microbatch gradient-accumulation pattern as the
formal recipe, using DDP ``no_sync`` for the first seven microbatches.

It intentionally exercises only the student VLM forward/backward.  The full
OPD smoke covers local frozen-teacher scoring; separating them makes a
collective failure attributable to the distributed gradient path.
"""

from __future__ import annotations

import argparse
import contextlib
import math
import os
import sys
import time
from datetime import timedelta
from pathlib import Path

import torch
import torch.distributed as dist
from PIL import Image
from torch.nn.parallel import DistributedDataParallel
from transformers import AutoProcessor, LlavaOnevisionForConditionalGeneration
from transformers.models.llava_onevision.modeling_llava_onevision import image_size_to_num_patches

PROJECT_ROOT = Path(__file__).resolve().parents[1]
if str(PROJECT_ROOT) not in sys.path:
    sys.path.insert(0, str(PROJECT_ROOT))

from config.loader import load_config
from data_utils.chart.data_collector import prepare_chart_rl_data
from data_utils.chart.teacher_gate import filter_chartqa_rows_with_teacher_gate
from data_utils.commom_util import collate_fn
from opsd_utils.deepspeed_utils import student_forward_chunk_size
from opsd_utils.teacher_batching import student_batch_num_images_tensor


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--config", default="config/config_opd_only_7b_chartqa.yaml")
    parser.add_argument("--updates", type=int, default=2)
    parser.add_argument("--gradient-accumulation-steps", type=int, default=8)
    parser.add_argument("--samples-per-rank", type=int, default=2)
    parser.add_argument(
        "--student-forward-chunk-size",
        type=int,
        default=0,
        help=(
            "Override the local student-forward chunk size. Zero uses the production "
            "student_forward_chunk_size() policy."
        ),
    )
    parser.add_argument("--sync-each-chunk", action="store_true")
    parser.add_argument("--trace-chunks", action="store_true")
    parser.add_argument("--process-group-timeout-seconds", type=int, default=120)
    return parser.parse_args()


def _memory_snapshot(device: torch.device) -> str:
    free_bytes, total_bytes = torch.cuda.mem_get_info(device)
    gib = 1024**3
    return (
        f"allocated_gib={torch.cuda.memory_allocated(device) / gib:.3f} "
        f"reserved_gib={torch.cuda.memory_reserved(device) / gib:.3f} "
        f"max_allocated_gib={torch.cuda.max_memory_allocated(device) / gib:.3f} "
        f"free_gib={free_bytes / gib:.3f} total_gib={total_bytes / gib:.3f}"
    )


def _rank_rows_with_extreme_shapes(
    config: dict,
    *,
    rank: int,
    world_size: int,
    samples_per_rank: int,
) -> tuple[list[dict], list[int]]:
    """Give rank 0 the largest images and all other ranks the smallest ones."""
    rows = prepare_chart_rl_data(config["dataset"]["train_dataset"])
    rows, _ = filter_chartqa_rows_with_teacher_gate(rows, config["dataset"]["teacher_gate"])
    processor = AutoProcessor.from_pretrained(
        config["model"]["pretrained_model_path"], local_files_only=True
    )
    grid = processor.image_processor.image_grid_pinpoints
    patch_size = processor.image_processor.size["height"]
    ranked: list[tuple[int, int, dict]] = []
    for index, row in enumerate(rows):
        with Image.open(row["image"]) as image:
            patches = image_size_to_num_patches(image.size, grid, patch_size)
        ranked.append((int(patches), index, row))
    ranked.sort(key=lambda item: (item[0], item[1]))

    if rank == 0:
        chosen = list(reversed(ranked[-samples_per_rank:]))
    else:
        start = (rank - 1) * samples_per_rank
        chosen = ranked[start : start + samples_per_rank]
    if len(chosen) != samples_per_rank:
        raise RuntimeError(
            "not enough ChartQA rows for distributed variable-shape DDP diagnostic"
        )
    return [row for _patches, _index, row in chosen], [patches for patches, _index, _row in chosen]


def _assert_extreme_shape_assignment(
    local_patch_counts: list[int], device: torch.device, rank: int, world_size: int) -> None:
    """Fail closed if the test no longer creates the intended rank skew."""
    local_max = torch.tensor([max(local_patch_counts)], device=device, dtype=torch.long)
    gathered = [torch.zeros_like(local_max) for _ in range(world_size)]
    dist.all_gather(gathered, local_max)
    maxima = [int(item.item()) for item in gathered]
    if rank == 0:
        print(f"DDP variable-shape assignment: per_rank_max_patches={maxima}", flush=True)
    if maxima[0] <= max(maxima[1:]):
        raise RuntimeError(
            "rank 0 did not receive a strictly larger image than all other ranks: "
            f"{maxima}"
        )


def main() -> None:
    args = parse_args()
    if args.updates <= 0 or args.gradient_accumulation_steps <= 0 or args.samples_per_rank <= 0:
        raise ValueError("--updates, --gradient-accumulation-steps, and --samples-per-rank must be positive")
    local_rank = int(os.environ["LOCAL_RANK"])
    torch.cuda.set_device(local_rank)
    if args.process_group_timeout_seconds <= 0:
        raise ValueError("--process-group-timeout-seconds must be positive")
    dist.init_process_group(
        backend="nccl",
        timeout=timedelta(seconds=args.process_group_timeout_seconds),
    )
    rank = dist.get_rank()
    world_size = dist.get_world_size()
    device = torch.device("cuda", local_rank)
    try:
        config = load_config(args.config)
        local_rows, local_patch_counts = _rank_rows_with_extreme_shapes(
            config,
            rank=rank,
            world_size=world_size,
            samples_per_rank=args.samples_per_rank,
        )
        _assert_extreme_shape_assignment(local_patch_counts, device, rank, world_size)

        model_path = config["model"]["pretrained_model_path"]
        processor = AutoProcessor.from_pretrained(model_path, local_files_only=True)
        processor.tokenizer.padding_side = "left"
        model = LlavaOnevisionForConditionalGeneration.from_pretrained(
            model_path,
            torch_dtype=torch.bfloat16,
            attn_implementation="flash_attention_2",
            local_files_only=True,
            low_cpu_mem_usage=True,
        )
        model.base_model.vision_tower.requires_grad_(False)
        model.config.use_cache = False
        model.to(device)
        ddp_model = DistributedDataParallel(
            model,
            device_ids=[local_rank],
            output_device=local_rank,
            find_unused_parameters=True,
        )
        optimizer = torch.optim.AdamW(
            (parameter for parameter in ddp_model.parameters() if parameter.requires_grad),
            lr=1.0e-6,
        )

        batch = collate_fn(local_rows, processor, label_id=None)
        input_ids = batch["input_ids"].to(device)
        attention_mask = batch["attention_mask"].to(device)
        pixel_values = batch["pixel_values"].to(device=device, dtype=torch.bfloat16)
        image_sizes = batch["image_sizes"].to(device)
        batch_num_images = student_batch_num_images_tensor(pixel_values, input_ids.size(0))
        production_chunk_size = student_forward_chunk_size(
            int(input_ids.size(0)), has_vision=True
        )
        chunk_size = int(args.student_forward_chunk_size or production_chunk_size)
        if chunk_size <= 0 or chunk_size > int(input_ids.size(0)):
            raise ValueError(
                "student forward chunk size must be within the local batch: "
                f"chunk_size={chunk_size}, batch_size={input_ids.size(0)}"
            )
        if rank == 0:
            print(
                "DDP variable-shape forward policy: "
                f"production_chunk_size={production_chunk_size} active_chunk_size={chunk_size} "
                f"forwards_per_backward={math.ceil(input_ids.size(0) / chunk_size)}",
                flush=True,
            )

        dist.barrier()
        started = time.perf_counter()
        for update in range(args.updates):
            optimizer.zero_grad(set_to_none=True)
            for micro_step in range(args.gradient_accumulation_steps):
                is_sync_step = micro_step == args.gradient_accumulation_steps - 1
                sync_context = contextlib.nullcontext() if is_sync_step else ddp_model.no_sync()
                with sync_context:
                    chunk_losses = []
                    for chunk_start in range(0, int(input_ids.size(0)), chunk_size):
                        chunk_end = min(chunk_start + chunk_size, int(input_ids.size(0)))
                        if args.trace_chunks:
                            print(
                                "DDP chunk enter: "
                                f"rank={rank}/{world_size} update={update} micro_step={micro_step} "
                                f"chunk={chunk_start}:{chunk_end} {_memory_snapshot(device)}",
                                flush=True,
                            )
                        outputs = ddp_model(
                            input_ids=input_ids[chunk_start:chunk_end],
                            attention_mask=attention_mask[chunk_start:chunk_end],
                            pixel_values=pixel_values[chunk_start:chunk_end],
                            image_sizes=image_sizes[chunk_start:chunk_end],
                            batch_num_images=batch_num_images[chunk_start:chunk_end],
                            logits_to_keep=2,
                        )
                        if args.sync_each_chunk:
                            torch.cuda.synchronize(device)
                        row_fraction = (chunk_end - chunk_start) / int(input_ids.size(0))
                        chunk_losses.append(outputs.logits.float().mean() * row_fraction)
                        if args.trace_chunks:
                            print(
                                "DDP chunk done: "
                                f"rank={rank}/{world_size} update={update} micro_step={micro_step} "
                                f"chunk={chunk_start}:{chunk_end} {_memory_snapshot(device)}",
                                flush=True,
                            )
                    loss = sum(chunk_losses) / args.gradient_accumulation_steps
                    loss.backward()
            optimizer.step()
            torch.cuda.synchronize(device)
            if rank == 0:
                print(
                    "DDP variable-shape update complete: "
                    f"update={update + 1}/{args.updates} loss={float(loss.detach().item()):.6f}",
                    flush=True,
                )
            dist.barrier()

        torch.cuda.synchronize(device)
        elapsed = time.perf_counter() - started
        print(
            "DDP VLM variable-shape stress passed: "
            f"rank={rank}/{world_size} patch_counts={local_patch_counts} "
            f"input_shape={tuple(input_ids.shape)} updates={args.updates} "
            f"ga={args.gradient_accumulation_steps} elapsed_s={elapsed:.3f}",
            flush=True,
        )
        dist.barrier()
    finally:
        if dist.is_initialized():
            dist.destroy_process_group()


if __name__ == "__main__":
    main()