§11 · Frameworks — DeepEP-V2 serving and training support on EFA measured

DeepEP-V2 on AWS EFA has been measured and validated across three serving frameworks and two training stacks, covering inference, full training loops, and RL post-training. Every framework passes AIPerf smoke tests and produces coherent output. This page consolidates the framework support matrix with measured numbers — all traceable to committed artifacts.

3 / 3

serving frameworks (pass AIPerf)

2 / 2

training stacks (Megatron + NeMo-RL)

0 errors

over 630 AIPerf requests total

Serving frameworks — AIPerf validated

3 frameworks serving DeepEP-V2 on EFA — all pass AIPerf 0.10.0 smoke (0 errors over ≥200 requests each)
framework	status	notes
SGLang	PASS measured	0 errors / 210 requests (§08)
vLLM	PASS measured	0 errors / 210 requests, EP_NUM_COMMS=8 doubles throughput (§09)
TRT-LLM	PASS measured	0 errors / 210 requests, native seam (§09)

source: sections §08 (V1 NVSHMEM backend) and §09 (public repro). All runs on 2× p5en.48xlarge H200 with AWS EFA, Qwen3-30B-A3B-FP8 EP16.

Training frameworks — loss-match and E2E validation

2 training stacks — full forward+backward+optimizer loops, loss convergence validated
framework	status	notes
Megatron-LM	PASS measured	digit-for-digit loss-match vs NCCL-alltoall baseline (§08)
NeMo-RL (GRPO)	PASS measured	full GRPO post-training loop, E2E coherent (§08)

source: section §08 (V1 NVSHMEM backend). Megatron-LM: 3-step train loss convergence, V1-vs-V2 A/B shows numerical equivalence. NeMo-RL: full policy+value+reward pipeline over DeepEP transport.

What this proves — and what it doesn't

it IS

proof DeepEP-V2 is a working drop-in across 3 serving engines and 2 training stacks on AWS EFA
AIPerf-validated for all serving frameworks (0 errors over 200+ requests each)
training-validated with loss-match (Megatron) and E2E RL (NeMo-RL GRPO)

it is NOT

a serving throughput win — see §04 for the honest DeepEP/dense A/B (0.65–0.97×)
an integration-cost claim — some frameworks needed patches (see §08 and §09 for details)

Campaigns 2026-06-25 → 2026-07-17 · 2× p5en.48xlarge (H200, 16 EFA NICs/node) · EP16 · all numbers trace to sections §08 and §09. Framework support: serving (SGLang, vLLM, TRT-LLM), training (Megatron-LM, NeMo-RL). Transport: NCCL-GIN proxy (NCCL_GIN_TYPE=2) and NVSHMEM/libfabric (V1 backend). Sanitized — no internal IPs, registry IDs, or hostnames.