|
1 | 1 | # MiniMax-H3-Image for text-to-image and image editing — Ascend NPU |
2 | 2 |
|
3 | | -> Ascend NPU deployment guide covering **Atlas 800T A3**, **Atlas 800I A2** |
| 3 | +> Ascend NPU deployment guide covering **Atlas 800I A2**, **Atlas 800T A3**, |
4 | 4 | > and **Atlas 350 (A5)**. It covers checkpoint preparation, the official |
5 | 5 | > image and container, deployment commands, request parameters, supported |
6 | | -> image sizes and supported features, plus the measured Atlas 350 |
7 | | -> performance. |
| 6 | +> image sizes and supported features, plus the measured Atlas 800I A2, |
| 7 | +> Atlas 800T A3 and Atlas 350 (A5) performance. |
8 | 8 |
|
9 | 9 | ## Summary |
10 | 10 |
|
@@ -142,12 +142,12 @@ matching the platform from the |
142 | 142 |
|
143 | 143 | | Platform | Image | |
144 | 144 | | --- | --- | |
145 | | -| Atlas 800I A2 | `quay.io/ascend/vllm-omni:v0.33.0rc1` | |
146 | | -| Atlas 800T A3 | `quay.io/ascend/vllm-omni:v0.33.0rc1-a3` | |
147 | | -| Atlas 350 (A5) | `quay.io/ascend/vllm-omni:v0.33.0rc1-a5` | |
| 145 | +| Atlas 800I A2 | `quay.io/ascend/vllm-omni:minimax-h3-image` | |
| 146 | +| Atlas 800T A3 | `quay.io/ascend/vllm-omni:minimax-h3-image-a3` | |
| 147 | +| Atlas 350 (A5) | `quay.io/ascend/vllm-omni:minimax-h3-image-a5` | |
148 | 148 |
|
149 | 149 | ```bash |
150 | | -export IMAGE=quay.io/ascend/vllm-omni:v0.33.0rc1-a3 |
| 150 | +export IMAGE=quay.io/ascend/vllm-omni:minimax-h3-image-a3 |
151 | 151 | export CONTAINER_NAME=h3-image |
152 | 152 |
|
153 | 153 | docker run -it -u root --name ${CONTAINER_NAME} \ |
@@ -372,6 +372,33 @@ reference, structured prompt and CPU generator with seed 1234. |
372 | 372 | | 2 NPUs, Ulysses SP2, text encoder TP2, DiT DLO with AllGather | 6.45 s | 6.45 s | 7.27 s | 18.89 s | 39.65 GiB | |
373 | 373 | | 1 NPU, DiT layer offload | 7.79 s | 7.76 s | 8.36 s | 31.10 s | 64.20 GiB | |
374 | 374 |
|
| 375 | +### Atlas 800T A3 performance |
| 376 | + |
| 377 | +Verified on Atlas 800T A3 (128 GiB per card as 2 × 64 GiB dies). The 1- and |
| 378 | +2-die profiles ran on one card, 4 dies on two cards, and 8 dies on four |
| 379 | +cards. The test case is the same as the Atlas 350 table: BF16, eager, six |
| 380 | +denoising steps, concurrency 1. Image: |
| 381 | +`quay.io/ascend/vllm-omni:minimax-h3-image-a3` (vLLM-Omni `eb8b991`). |
| 382 | + |
| 383 | +The 8-die profile uses DiT TP8 and text-encoder TP8 with all weights |
| 384 | +resident. The 4-die profile uses Ulysses SP4, text-encoder TP4 and DiT |
| 385 | +distributed layerwise offload with AllGather: keeping every weight resident |
| 386 | +does not fit in 64 GiB per die. The 2-die profile uses Ulysses SP2, |
| 387 | +text-encoder TP2 (the encoder stays resident on the dies) and the same DiT |
| 388 | +offload. The 1-die profile offloads both DiT and text encoder to host |
| 389 | +memory, so it holds much less HBM per die. |
| 390 | + |
| 391 | +The 4- and 8-die runs shared their cards with idle neighbour processes that |
| 392 | +held up to 10 GiB per die at 0% AICore. The reported peak excludes that |
| 393 | +baseline, and the timings were taken while those processes stayed idle. |
| 394 | + |
| 395 | +| Profile | T2I 1024² | T2I 2048² | Edit 2048² | Edit 4096² | Peak HBM / die | |
| 396 | +| --- | ---: | ---: | ---: | ---: | ---: | |
| 397 | +| 8 dies (4 cards), DiT TP8, text encoder TP8, fully resident | 2.12 s | 2.10 s | 3.36 s | 9.65 s | 19.45 GiB | |
| 398 | +| 4 dies (2 cards), Ulysses SP4, text encoder TP4, DiT DLO with AllGather | 4.48 s | 4.47 s | 5.31 s | 14.87 s | 20.32 GiB | |
| 399 | +| 2 dies, Ulysses SP2, text encoder TP2, DiT DLO with AllGather | 6.30 s | 6.10 s | 8.71 s | 28.09 s | 36.20 GiB | |
| 400 | +| 1 die, DiT + text encoder layer offload | 10.86 s | 10.79 s | 14.87 s | 54.59 s | 12.26 GiB | |
| 401 | + |
375 | 402 | ### Atlas 800I A2 performance |
376 | 403 |
|
377 | 404 | The test case is the same as the Atlas 350 table. |
|
0 commit comments