2nd/10_Wiki/Topics/Architecture/Parallel-Computing-in-AI.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-parallel-computing-in-ai	Parallel Computing in AI	10_Wiki/Topics	verified	self	AI parallelism distributed AI training AI parallel computing	none	A	0.9	applied	parallelism ai distributed gpu training		2026-05-10	pending	language framework python pytorch

Parallel Computing in AI

매 한 줄

"매 AI 의 parallel 은 단순 distributed 가 아니라 4축 (Data, Tensor, Pipeline, Expert) × 2 모드 (training/inference) 의 조합 problem". 매 2026 Llama 3 405B 학습 = 16384 GPU, GPT-5 추론 = 매 expert sharding + speculative decoding. 매 parallel strategy 의 mismatch = 매 ROI 폭락.

매 핵심

매 4 축

1. Data Parallel (DP): 매 batch 분할, 매 weight 동기.
2. Tensor Parallel (TP): 매 single layer 의 weight matrix 분할.
3. Pipeline Parallel (PP): 매 layer stack 분할.
4. Expert Parallel (EP): 매 MoE 의 expert 분할.

매 + 보조 axis

• Sequence Parallel (SP): 매 long context 의 token 차원 분할 (Ring Attention, Ulysses).
• Context Parallel (CP): 매 attention 의 KV partial.
• ZeRO sharding (DP variant): 매 optimizer/grad/param 분할.

매 training vs inference

측면	Training	Inference
주축	DP + ZeRO	TP + EP
Batch	큰 micro-batch	매 1 user → 매 small
Memory	activation 중심	KV cache 중심
Comm	all-reduce (grad)	all-reduce (TP), all-to-all (MoE)
도구	DeepSpeed, Megatron, FSDP	vLLM, TensorRT-LLM, SGLang

매 hardware constraint

• NVLink (intra-node, 900GB/s): 매 TP 친화.
• InfiniBand / NVLink Switch (inter-node, 400Gbps NDR): 매 PP/DP.
• HBM3e (192GB on B200): 매 더 큰 model fit + 매 sharding 절약.

💻 패턴

FSDP2 (PyTorch 2.6+, 2026)

import torch
from torch.distributed.fsdp import fully_shard, MixedPrecisionPolicy

model = build_transformer()
mp = MixedPrecisionPolicy(param_dtype=torch.bfloat16, reduce_dtype=torch.float32)

for layer in model.layers:
    fully_shard(layer, mp_policy=mp)
fully_shard(model, mp_policy=mp)

opt = torch.optim.AdamW(model.parameters(), lr=3e-4, fused=True)
for batch in loader:
    loss = model(batch).loss
    loss.backward()
    opt.step(); opt.zero_grad()

TP with DTensor / device_mesh

from torch.distributed.tensor import DTensor, Shard, Replicate
from torch.distributed.device_mesh import init_device_mesh

mesh = init_device_mesh("cuda", (DP, TP), mesh_dim_names=("dp","tp"))

# Column-parallel linear (매 weight col-sharded)
w = nn.Parameter(torch.empty(out, in_))
w_dt = DTensor.from_local(w, mesh["tp"], [Shard(0)])
# 매 forward: x @ w_dt.T → 매 partial output, all-reduce 자동

MoE expert parallel (DeepSpeed-MoE / Megatron)

class MoELayer(nn.Module):
    def __init__(self, experts, top_k=2):
        super().__init__()
        self.experts = nn.ModuleList(experts)  # 매 sharded across EP group
        self.gate = nn.Linear(d, len(experts))
        self.k = top_k
    def forward(self, x):
        logits = self.gate(x)
        topk = logits.topk(self.k, dim=-1)
        # 매 all-to-all dispatch tokens to expert-owning rank
        dispatched = all_to_all_dispatch(x, topk.indices, ep_group)
        out = local_experts(dispatched)
        return all_to_all_combine(out, topk.indices, topk.values, ep_group)

vLLM tensor-parallel inference

from vllm import LLM, SamplingParams

llm = LLM(
    model="meta-llama/Llama-3.3-70B-Instruct",
    tensor_parallel_size=4,        # 매 4 GPU TP
    pipeline_parallel_size=1,
    gpu_memory_utilization=0.92,
    enable_chunked_prefill=True,
    enable_prefix_caching=True,
)
out = llm.generate(["매 안녕"], SamplingParams(temperature=0.7, max_tokens=128))

Ring Attention (sequence parallel for 1M context)

def ring_attention(q, k, v, sp_group):
    # 매 each rank holds 1/N of sequence
    out = torch.zeros_like(q); lse = torch.full(...)
    k_blk, v_blk = k, v
    for step in range(world_size(sp_group)):
        partial, l = flash_attn_partial(q, k_blk, v_blk)
        out, lse = log_sum_exp_combine(out, lse, partial, l)
        k_blk, v_blk = ring_send_recv(k_blk, v_blk, sp_group)
    return out

3D parallel mesh (Megatron-Core 2026)

mesh = init_device_mesh(
    "cuda", (DP_replica, DP_shard, PP, TP),
    mesh_dim_names=("dp_r","dp_s","pp","tp"),
)
# Llama 3 405B: dp_r=16, dp_s=8, pp=16, tp=8 → 16384 GPUs

매 결정 기준

상황	Approach
≤ 8B model, 1 node	DP + ZeRO-2
70B, multi-node	FSDP2 (DP shard) + TP intra-node
405B+	TP × PP × DP (3D)
MoE (Mixtral, GPT-5 style)	+ EP (all-to-all)
1M+ context	+ SP (Ring/Ulysses)
Inference 1 user	TP only, no DP
Inference batch	TP + continuous batching (vLLM)
MoE inference	EP + speculative decoding

기본값: 매 training = FSDP2 + TP, 매 inference = vLLM TP.

🔗 Graph

• 부모: Distributed Computing · High Performance Computing
• 변형: Data Parallelism · Tensor Parallelism · Pipeline Parallelism · Expert Parallelism
• 응용: LLM Training · vLLM · DeepSpeed · Megatron-LM
• Adjacent: ZeRO Optimizer · FSDP · Ring Attention · Mixture of Experts

🤖 LLM 활용

언제: 매 model 또는 batch 가 single-GPU mem 초과. 매 throughput / latency SLO 의 multi-GPU 필요. 언제 X: 매 single GPU fits + 매 batch latency 만족 — 매 multi-GPU overhead 가 손해.

❌ 안티패턴

• TP across nodes: 매 IB 의 NVLink 대비 5-10x 느림 → 매 stall.
• PP without DP: 매 batch 의 micro-batch 한계 → 매 throughput cap.
• MoE EP without all-to-all opt: 매 NCCL all-to-all 의 성능 ↓ → 매 GroupedGEMM kernel 필요.
• Sequence parallel without flash-attn: 매 attention recompute 폭증.
• Mixed precision without loss scaling (FP16): 매 underflow → 매 loss NaN. → BF16 권장.

🧪 검증 / 중복

• Verified (Megatron-LM paper, FSDP2 docs 2026, vLLM docs, DeepSpeed-MoE paper).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — 4-axis parallelism + training/inference split