[CI][ROCm] Mirror large-model GSM8K evaluations on MI355 (#58683)
Signed-off-by: Andreas Karatzas <andreas.karatzas@protonmail.com> Co-authored-by: Codex <noreply@openai.com>
Andreas Karatzas committed
Sep 29, 2026 at 02:06 UTC
0ae480ff1a1b07a6959c3b19113739ffd2b60820
14 files changed
+446
-5
.buildkite/test-amd.yaml
+165
@@ -4248,6 +4248,171 @@ steps:
4248
- export VLLM_WORKER_MULTIPROC_METHOD=spawn
4249
- pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-gfx950-large.txt
4250
4251
+- label: ":amd: (MI355) LM Eval DeepSeek R1 TP"
4252
+ timeout_in_minutes: 65
4253
+ mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
4254
+ dind: false
4255
+ agent_pool: mi355_4
4256
+ optional: true
4257
+ num_gpus: 4
4258
+ working_dir: "/vllm-workspace/tests"
4259
+ source_file_dependencies:
4260
+ - .buildkite/test_areas/lm_eval.yaml
4261
+ - tests/evals/gsm8k/
4262
+ - tests/utils.py
4263
+ - tests/conftest.py
4264
+ - vllm/config/
4265
+ - vllm/distributed/
4266
+ - vllm/model_executor/models/
4267
+ - vllm/models/
4268
+ - vllm/model_executor/model_loader/
4269
+ - vllm/model_executor/layers/quantization/
4270
+ - vllm/model_executor/layers/fused_moe/
4271
+ - vllm/v1/attention/
4272
+ - vllm/v1/spec_decode/
4273
+ - vllm/v1/worker/gpu/spec_decode/
4274
+ - vllm/v1/core/kv_cache_utils.py
4275
+ - vllm/model_executor/layers/layernorm.py
4276
+ - csrc/
4277
+ - vllm/_aiter_ops.py
4278
+ - vllm/platforms/rocm.py
4279
+ - vllm/model_executor/layers/sparse_attn_indexer.py
4280
+ commands:
4281
+ - export VLLM_WORKER_MULTIPROC_METHOD=spawn
4282
+ - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi355-deepseek-r1-tp.txt
4283
+
4284
+- label: ":amd: (MI355) LM Eval DeepSeek R1 DP"
4285
+ timeout_in_minutes: 65
4286
+ mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
4287
+ dind: false
4288
+ agent_pool: mi355_4
4289
+ optional: true
4290
+ num_gpus: 4
4291
+ working_dir: "/vllm-workspace/tests"
4292
+ source_file_dependencies:
4293
+ - .buildkite/test_areas/lm_eval.yaml
4294
+ - tests/evals/gsm8k/
4295
+ - tests/utils.py
4296
+ - tests/conftest.py
4297
+ - vllm/config/
4298
+ - vllm/distributed/
4299
+ - vllm/model_executor/models/
4300
+ - vllm/models/
4301
+ - vllm/model_executor/model_loader/
4302
+ - vllm/model_executor/layers/quantization/
4303
+ - vllm/model_executor/layers/fused_moe/
4304
+ - vllm/v1/attention/
4305
+ - vllm/v1/spec_decode/
4306
+ - vllm/v1/worker/gpu/spec_decode/
4307
+ - vllm/v1/core/kv_cache_utils.py
4308
+ - vllm/model_executor/layers/layernorm.py
4309
+ - csrc/
4310
+ - vllm/_aiter_ops.py
4311
+ - vllm/platforms/rocm.py
4312
+ - vllm/model_executor/layers/sparse_attn_indexer.py
4313
+ commands:
4314
+ - export VLLM_WORKER_MULTIPROC_METHOD=spawn
4315
+ - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi355-deepseek-r1-dp.txt
4316
+
4317
+- label: ":amd: (MI355) LM Eval DeepSeek V3.2 TP"
4318
+ timeout_in_minutes: 65
4319
+ mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
4320
+ dind: false
4321
+ agent_pool: mi355_4
4322
+ optional: true
4323
+ num_gpus: 4
4324
+ working_dir: "/vllm-workspace/tests"
4325
+ source_file_dependencies:
4326
+ - .buildkite/test_areas/lm_eval.yaml
4327
+ - tests/evals/gsm8k/
4328
+ - tests/utils.py
4329
+ - tests/conftest.py
4330
+ - vllm/config/
4331
+ - vllm/distributed/
4332
+ - vllm/model_executor/models/
4333
+ - vllm/models/
4334
+ - vllm/model_executor/model_loader/
4335
+ - vllm/model_executor/layers/quantization/
4336
+ - vllm/model_executor/layers/fused_moe/
4337
+ - vllm/v1/attention/
4338
+ - vllm/v1/spec_decode/
4339
+ - vllm/v1/worker/gpu/spec_decode/
4340
+ - vllm/v1/core/kv_cache_utils.py
4341
+ - vllm/model_executor/layers/layernorm.py
4342
+ - csrc/
4343
+ - vllm/_aiter_ops.py
4344
+ - vllm/platforms/rocm.py
4345
+ - vllm/model_executor/layers/sparse_attn_indexer.py
4346
+ commands:
4347
+ - export VLLM_WORKER_MULTIPROC_METHOD=spawn
4348
+ - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi355-deepseek-v32-tp.txt
4349
+
4350
+- label: ":amd: (MI355) LM Eval DeepSeek V3.2 DP"
4351
+ timeout_in_minutes: 65
4352
+ mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
4353
+ dind: false
4354
+ agent_pool: mi355_4
4355
+ optional: true
4356
+ num_gpus: 4
4357
+ working_dir: "/vllm-workspace/tests"
4358
+ source_file_dependencies:
4359
+ - .buildkite/test_areas/lm_eval.yaml
4360
+ - tests/evals/gsm8k/
4361
+ - tests/utils.py
4362
+ - tests/conftest.py
4363
+ - vllm/config/
4364
+ - vllm/distributed/
4365
+ - vllm/model_executor/models/
4366
+ - vllm/models/
4367
+ - vllm/model_executor/model_loader/
4368
+ - vllm/model_executor/layers/quantization/
4369
+ - vllm/model_executor/layers/fused_moe/
4370
+ - vllm/v1/attention/
4371
+ - vllm/v1/spec_decode/
4372
+ - vllm/v1/worker/gpu/spec_decode/
4373
+ - vllm/v1/core/kv_cache_utils.py
4374
+ - vllm/model_executor/layers/layernorm.py
4375
+ - csrc/
4376
+ - vllm/_aiter_ops.py
4377
+ - vllm/platforms/rocm.py
4378
+ - vllm/model_executor/layers/sparse_attn_indexer.py
4379
+ commands:
4380
+ - export VLLM_WORKER_MULTIPROC_METHOD=spawn
4381
+ - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi355-deepseek-v32-dp.txt
4382
+
4383
+- label: ":amd: (MI355) LM Eval Nemotron 3 Super"
4384
+ timeout_in_minutes: 65
4385
+ mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
4386
+ dind: false
4387
+ agent_pool: mi355_4
4388
+ optional: true
4389
+ num_gpus: 4
4390
+ working_dir: "/vllm-workspace/tests"
4391
+ source_file_dependencies:
4392
+ - .buildkite/test_areas/lm_eval.yaml
4393
+ - tests/evals/gsm8k/
4394
+ - tests/utils.py
4395
+ - tests/conftest.py
4396
+ - vllm/config/
4397
+ - vllm/distributed/
4398
+ - vllm/model_executor/models/
4399
+ - vllm/models/
4400
+ - vllm/model_executor/model_loader/
4401
+ - vllm/model_executor/layers/quantization/
4402
+ - vllm/model_executor/layers/fused_moe/
4403
+ - vllm/v1/attention/
4404
+ - vllm/v1/spec_decode/
4405
+ - vllm/v1/worker/gpu/spec_decode/
4406
+ - vllm/v1/core/kv_cache_utils.py
4407
+ - vllm/model_executor/layers/layernorm.py
4408
+ - csrc/
4409
+ - vllm/_aiter_ops.py
4410
+ - vllm/platforms/rocm.py
4411
+ - vllm/model_executor/layers/mamba/
4412
+ commands:
4413
+ - export VLLM_WORKER_MULTIPROC_METHOD=spawn
4414
+ - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi355-nemotron-3-super.txt
4415
+
4416
#--------------------------------------------------------- mi355 · examples ----------------------------------------------------------#
4417
4418
- label: ":amd: (MI355 DPX) Examples" # TBD
.buildkite/test_areas/lm_eval.yaml
+170
@@ -436,6 +436,40 @@ steps:
436
num_devices: 8
437
commands:
438
- pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-h200-deepseek-r1-tp.txt
439
+ mirror:
440
+ amd:
441
+ label: ":amd: (MI355) LM Eval DeepSeek R1 TP"
442
+ dind: false
443
+ device: mi355_4
444
+ num_devices: 4
445
+ timeout_in_minutes: 65
446
+ working_dir: /vllm-workspace/tests
447
+ depends_on:
448
+ - image-build-amd
449
+ source_file_dependencies:
450
+ - .buildkite/test_areas/lm_eval.yaml
451
+ - tests/evals/gsm8k/
452
+ - tests/utils.py
453
+ - tests/conftest.py
454
+ - vllm/config/
455
+ - vllm/distributed/
456
+ - vllm/model_executor/models/
457
+ - vllm/models/
458
+ - vllm/model_executor/model_loader/
459
+ - vllm/model_executor/layers/quantization/
460
+ - vllm/model_executor/layers/fused_moe/
461
+ - vllm/v1/attention/
462
+ - vllm/v1/spec_decode/
463
+ - vllm/v1/worker/gpu/spec_decode/
464
+ - vllm/v1/core/kv_cache_utils.py
465
+ - vllm/model_executor/layers/layernorm.py
466
+ - csrc/
467
+ - vllm/_aiter_ops.py
468
+ - vllm/platforms/rocm.py
469
+ - vllm/model_executor/layers/sparse_attn_indexer.py
470
+ commands:
471
+ - export VLLM_WORKER_MULTIPROC_METHOD=spawn
472
+ - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi355-deepseek-r1-tp.txt
473
474
- label: ":nvidia: (H200) LM Eval DeepSeek R1 DP"
475
key: lm-eval-deepseek-r1-dp-8xh200
@@ -445,6 +479,40 @@ steps:
479
num_devices: 8
480
commands:
481
- pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-h200-deepseek-r1-dp.txt
482
+ mirror:
483
+ amd:
484
+ label: ":amd: (MI355) LM Eval DeepSeek R1 DP"
485
+ dind: false
486
+ device: mi355_4
487
+ num_devices: 4
488
+ timeout_in_minutes: 65
489
+ working_dir: /vllm-workspace/tests
490
+ depends_on:
491
+ - image-build-amd
492
+ source_file_dependencies:
493
+ - .buildkite/test_areas/lm_eval.yaml
494
+ - tests/evals/gsm8k/
495
+ - tests/utils.py
496
+ - tests/conftest.py
497
+ - vllm/config/
498
+ - vllm/distributed/
499
+ - vllm/model_executor/models/
500
+ - vllm/models/
501
+ - vllm/model_executor/model_loader/
502
+ - vllm/model_executor/layers/quantization/
503
+ - vllm/model_executor/layers/fused_moe/
504
+ - vllm/v1/attention/
505
+ - vllm/v1/spec_decode/
506
+ - vllm/v1/worker/gpu/spec_decode/
507
+ - vllm/v1/core/kv_cache_utils.py
508
+ - vllm/model_executor/layers/layernorm.py
509
+ - csrc/
510
+ - vllm/_aiter_ops.py
511
+ - vllm/platforms/rocm.py
512
+ - vllm/model_executor/layers/sparse_attn_indexer.py
513
+ commands:
514
+ - export VLLM_WORKER_MULTIPROC_METHOD=spawn
515
+ - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi355-deepseek-r1-dp.txt
516
517
- label: ":nvidia: (H200) LM Eval DeepSeek V3.2 TP"
518
key: lm-eval-deepseek-v32-tp-8xh200
@@ -454,6 +522,40 @@ steps:
522
num_devices: 8
523
commands:
524
- pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-h200-deepseek-v32-tp.txt
525
+ mirror:
526
+ amd:
527
+ label: ":amd: (MI355) LM Eval DeepSeek V3.2 TP"
528
+ dind: false
529
+ device: mi355_4
530
+ num_devices: 4
531
+ timeout_in_minutes: 65
532
+ working_dir: /vllm-workspace/tests
533
+ depends_on:
534
+ - image-build-amd
535
+ source_file_dependencies:
536
+ - .buildkite/test_areas/lm_eval.yaml
537
+ - tests/evals/gsm8k/
538
+ - tests/utils.py
539
+ - tests/conftest.py
540
+ - vllm/config/
541
+ - vllm/distributed/
542
+ - vllm/model_executor/models/
543
+ - vllm/models/
544
+ - vllm/model_executor/model_loader/
545
+ - vllm/model_executor/layers/quantization/
546
+ - vllm/model_executor/layers/fused_moe/
547
+ - vllm/v1/attention/
548
+ - vllm/v1/spec_decode/
549
+ - vllm/v1/worker/gpu/spec_decode/
550
+ - vllm/v1/core/kv_cache_utils.py
551
+ - vllm/model_executor/layers/layernorm.py
552
+ - csrc/
553
+ - vllm/_aiter_ops.py
554
+ - vllm/platforms/rocm.py
555
+ - vllm/model_executor/layers/sparse_attn_indexer.py
556
+ commands:
557
+ - export VLLM_WORKER_MULTIPROC_METHOD=spawn
558
+ - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi355-deepseek-v32-tp.txt
559
560
- label: ":nvidia: (H200) LM Eval DeepSeek V3.2 DP"
561
key: lm-eval-deepseek-v32-dp-8xh200
@@ -463,6 +565,40 @@ steps:
565
num_devices: 8
566
commands:
567
- pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-h200-deepseek-v32-dp.txt
568
+ mirror:
569
+ amd:
570
+ label: ":amd: (MI355) LM Eval DeepSeek V3.2 DP"
571
+ dind: false
572
+ device: mi355_4
573
+ num_devices: 4
574
+ timeout_in_minutes: 65
575
+ working_dir: /vllm-workspace/tests
576
+ depends_on:
577
+ - image-build-amd
578
+ source_file_dependencies:
579
+ - .buildkite/test_areas/lm_eval.yaml
580
+ - tests/evals/gsm8k/
581
+ - tests/utils.py
582
+ - tests/conftest.py
583
+ - vllm/config/
584
+ - vllm/distributed/
585
+ - vllm/model_executor/models/
586
+ - vllm/models/
587
+ - vllm/model_executor/model_loader/
588
+ - vllm/model_executor/layers/quantization/
589
+ - vllm/model_executor/layers/fused_moe/
590
+ - vllm/v1/attention/
591
+ - vllm/v1/spec_decode/
592
+ - vllm/v1/worker/gpu/spec_decode/
593
+ - vllm/v1/core/kv_cache_utils.py
594
+ - vllm/model_executor/layers/layernorm.py
595
+ - csrc/
596
+ - vllm/_aiter_ops.py
597
+ - vllm/platforms/rocm.py
598
+ - vllm/model_executor/layers/sparse_attn_indexer.py
599
+ commands:
600
+ - export VLLM_WORKER_MULTIPROC_METHOD=spawn
601
+ - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi355-deepseek-v32-dp.txt
602
603
- label: ":nvidia: (H200) LM Eval Nemotron 3 Super"
604
key: lm-eval-nemotron-3-super-8xh200
@@ -472,6 +608,40 @@ steps:
608
num_devices: 8
609
commands:
610
- pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-h200-nemotron-3-super.txt
611
+ mirror:
612
+ amd:
613
+ label: ":amd: (MI355) LM Eval Nemotron 3 Super"
614
+ dind: false
615
+ device: mi355_4
616
+ num_devices: 4
617
+ timeout_in_minutes: 65
618
+ working_dir: /vllm-workspace/tests
619
+ depends_on:
620
+ - image-build-amd
621
+ source_file_dependencies:
622
+ - .buildkite/test_areas/lm_eval.yaml
623
+ - tests/evals/gsm8k/
624
+ - tests/utils.py
625
+ - tests/conftest.py
626
+ - vllm/config/
627
+ - vllm/distributed/
628
+ - vllm/model_executor/models/
629
+ - vllm/models/
630
+ - vllm/model_executor/model_loader/
631
+ - vllm/model_executor/layers/quantization/
632
+ - vllm/model_executor/layers/fused_moe/
633
+ - vllm/v1/attention/
634
+ - vllm/v1/spec_decode/
635
+ - vllm/v1/worker/gpu/spec_decode/
636
+ - vllm/v1/core/kv_cache_utils.py
637
+ - vllm/model_executor/layers/layernorm.py
638
+ - csrc/
639
+ - vllm/_aiter_ops.py
640
+ - vllm/platforms/rocm.py
641
+ - vllm/model_executor/layers/mamba/
642
+ commands:
643
+ - export VLLM_WORKER_MULTIPROC_METHOD=spawn
644
+ - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi355-nemotron-3-super.txt
645
646
- label: ":amd: (MI300) LM Eval Large Models"
647
key: amd-lm-eval-large-models
tests/evals/gsm8k/configs/DeepSeek-R1-DP_MI355.yaml
new
+17
@@ -0,0 +1,17 @@
1
+model_name: "deepseek-ai/DeepSeek-R1"
2
+accuracy_threshold: 0.95
3
+num_questions: 1319
4
+num_fewshot: 5
5
+startup_max_wait_seconds: 1200
6
+env:
7
+ VLLM_ROCM_USE_AITER: "1"
8
+ VLLM_ROCM_USE_AITER_MLA: "1"
9
+ VLLM_ROCM_USE_AITER_MOE: "1"
10
+server_args: >-
11
+ --enforce-eager
12
+ --max-model-len 4096
13
+ --data-parallel-size 4
14
+ --enable-expert-parallel
15
+ --attention-backend ROCM_AITER_MLA
16
+ --moe-backend aiter
17
+ --speculative-config '{"method":"mtp","num_speculative_tokens":3}'
tests/evals/gsm8k/configs/DeepSeek-R1-TP_MI355.yaml
new
+17
@@ -0,0 +1,17 @@
1
+model_name: "deepseek-ai/DeepSeek-R1"
2
+accuracy_threshold: 0.95
3
+num_questions: 1319
4
+num_fewshot: 5
5
+startup_max_wait_seconds: 1200
6
+env:
7
+ VLLM_ROCM_USE_AITER: "1"
8
+ VLLM_ROCM_USE_AITER_MLA: "1"
9
+ VLLM_ROCM_USE_AITER_MOE: "1"
10
+server_args: >-
11
+ --enforce-eager
12
+ --max-model-len 4096
13
+ --tensor-parallel-size 4
14
+ --enable-expert-parallel
15
+ --attention-backend ROCM_AITER_MLA
16
+ --moe-backend aiter
17
+ --speculative-config '{"method":"mtp","num_speculative_tokens":3}'
tests/evals/gsm8k/configs/DeepSeek-V3.2-DP_MI355.yaml
new
+17
@@ -0,0 +1,17 @@
1
+model_name: "deepseek-ai/DeepSeek-V3.2"
2
+accuracy_threshold: 0.95
3
+num_questions: 1319
4
+num_fewshot: 5
5
+startup_max_wait_seconds: 1200
6
+env:
7
+ VLLM_ROCM_USE_AITER: "1"
8
+ VLLM_ROCM_USE_AITER_MLA: "1"
9
+ VLLM_ROCM_USE_AITER_MOE: "1"
10
+server_args: >-
11
+ --enforce-eager
12
+ --max-model-len 4096
13
+ --data-parallel-size 4
14
+ --enable-expert-parallel
15
+ --attention-backend ROCM_AITER_MLA_SPARSE
16
+ --moe-backend aiter
17
+ --speculative-config '{"method":"mtp","num_speculative_tokens":3}'
tests/evals/gsm8k/configs/DeepSeek-V3.2-TP_MI355.yaml
new
+17
@@ -0,0 +1,17 @@
1
+model_name: "deepseek-ai/DeepSeek-V3.2"
2
+accuracy_threshold: 0.95
3
+num_questions: 1319
4
+num_fewshot: 5
5
+startup_max_wait_seconds: 1200
6
+env:
7
+ VLLM_ROCM_USE_AITER: "1"
8
+ VLLM_ROCM_USE_AITER_MLA: "1"
9
+ VLLM_ROCM_USE_AITER_MOE: "1"
10
+server_args: >-
11
+ --enforce-eager
12
+ --max-model-len 4096
13
+ --tensor-parallel-size 4
14
+ --enable-expert-parallel
15
+ --attention-backend ROCM_AITER_MLA_SPARSE
16
+ --moe-backend aiter
17
+ --speculative-config '{"method":"mtp","num_speculative_tokens":3}'
tests/evals/gsm8k/configs/Nemotron-3-Super-120B-A12B-BF16_MI355.yaml
new
+15
@@ -0,0 +1,15 @@
1
+model_name: "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16"
2
+accuracy_threshold: 0.93
3
+num_questions: 1319
4
+num_fewshot: 5
5
+startup_max_wait_seconds: 1200
6
+env:
7
+ VLLM_ROCM_USE_AITER: "1"
8
+server_args: >-
9
+ --enforce-eager
10
+ --max-model-len 4096
11
+ --tensor-parallel-size 4
12
+ --enable-expert-parallel
13
+ --mamba-backend triton
14
+ --moe-backend triton
15
+ --speculative-config '{"method":"mtp","num_speculative_tokens":5}'
tests/evals/gsm8k/configs/Nemotron-3-Super-120B-A12B-FP8_MI355.yaml
new
+15
@@ -0,0 +1,15 @@
1
+model_name: "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8"
2
+accuracy_threshold: 0.93
3
+num_questions: 1319
4
+num_fewshot: 5
5
+startup_max_wait_seconds: 1200
6
+env:
7
+ VLLM_ROCM_USE_AITER: "1"
8
+server_args: >-
9
+ --enforce-eager
10
+ --max-model-len 4096
11
+ --tensor-parallel-size 4
12
+ --enable-expert-parallel
13
+ --mamba-backend triton
14
+ --moe-backend triton
15
+ --speculative-config '{"method":"mtp","num_speculative_tokens":5}'
tests/evals/gsm8k/configs/models-mi355-deepseek-r1-dp.txt
new
+1
@@ -0,0 +1 @@
1
+DeepSeek-R1-DP_MI355.yaml
tests/evals/gsm8k/configs/models-mi355-deepseek-r1-tp.txt
new
+1
@@ -0,0 +1 @@
1
+DeepSeek-R1-TP_MI355.yaml
tests/evals/gsm8k/configs/models-mi355-deepseek-v32-dp.txt
new
+1
@@ -0,0 +1 @@
1
+DeepSeek-V3.2-DP_MI355.yaml
tests/evals/gsm8k/configs/models-mi355-deepseek-v32-tp.txt
new
+1
@@ -0,0 +1 @@
1
+DeepSeek-V3.2-TP_MI355.yaml
tests/evals/gsm8k/configs/models-mi355-nemotron-3-super.txt
new
+2
@@ -0,0 +1,2 @@
1
+Nemotron-3-Super-120B-A12B-BF16_MI355.yaml
2
+Nemotron-3-Super-120B-A12B-FP8_MI355.yaml
tests/evals/gsm8k/test_gsm8k_correctness.py
+7
-5
@@ -109,15 +109,17 @@ def test_gsm8k_correctness(config_filename):
109
"Its W2A16 (uint2b2) scheme has no kernel outside CUDA."
110
)
111
112
- # TODO(akaratza): Enable DeepSeek-V3.2 and DeepSeek-R1 on ROCm platforms
112
if current_platform.is_rocm() and (
113
"deepseek-ai/DeepSeek-V3.2" in eval_config["model_name"]
114
or "deepseek-ai/DeepSeek-R1" in eval_config["model_name"]
115
):
117
- pytest.skip(
118
- "Skipping DeepSeek-V3.2 and DeepSeek-R1 on ROCm platforms "
119
- "due to agent pool disk space issues and pod evictions."
120
- )
116
+ from vllm.platforms.rocm import get_cdna_version
117
+
118
+ if get_cdna_version() < 4:
119
+ pytest.skip(
120
+ "Skipping DeepSeek-V3.2 and DeepSeek-R1 on ROCm platforms below "
121
+ "CDNA 4 due to agent pool disk space issues and pod evictions."
122
+ )
123
if current_platform.is_rocm() and (
124
"Qwen3.5-35B-A3B-MXFP4-AITER-TP2" in config_filename.name
125
):