diff --git a/tests/integration/defs/accuracy/references/gsm8k.yaml b/tests/integration/defs/accuracy/references/gsm8k.yaml
index 6f77797a8069..15d0ac772cfa 100644
--- a/tests/integration/defs/accuracy/references/gsm8k.yaml
+++ b/tests/integration/defs/accuracy/references/gsm8k.yaml
@@ -515,22 +515,3 @@ zai-org/GLM-5.2:
kv_cache_quant_algo: FP8
spec_dec_algo: MTP
accuracy: 90
-# Step-3.7-Flash text decoder (MoE) GSM8K, full 1319-sample split, TP4/EP4 with
-# TRTLLM attention + MoE backends. FP8 measured on the FP8 block-scale
-# checkpoint; NVFP4 measured on the modelopt NVFP4 export (FP8 KV cache). MTP
-# (mtp_nextn=3) is lossless for both FP8 and NVFP4 and reuses the non-spec
-# baseline accuracy.
-stepfun-ai/Step-3.7-Flash:
- - accuracy: 88
- - quant_algo: FP8_BLOCK_SCALES
- accuracy: 88
- - quant_algo: FP8_BLOCK_SCALES
- spec_dec_algo: MTP
- accuracy: 88
- - quant_algo: NVFP4
- kv_cache_quant_algo: FP8
- accuracy: 88
- - quant_algo: NVFP4
- kv_cache_quant_algo: FP8
- spec_dec_algo: MTP
- accuracy: 88
diff --git a/tests/integration/defs/accuracy/references/mmmu.yaml b/tests/integration/defs/accuracy/references/mmmu.yaml
index 17c1469d9ced..480df410b30f 100644
--- a/tests/integration/defs/accuracy/references/mmmu.yaml
+++ b/tests/integration/defs/accuracy/references/mmmu.yaml
@@ -86,22 +86,3 @@ moonshotai/Kimi-K2.5:
- quant_algo: NVFP4
kv_cache_quant_algo: FP8
accuracy: 81.56
-# Step-3.7-Flash multimodal (PerceptionEncoder vision tower + MoE text decoder).
-# Reasoning model: ... traces are stripped before MMMU answer
-# extraction (see test_llm_api_pytorch_multimodal.py::TestStep3_7). FP8 measured
-# on the FP8 block-scale checkpoint over the full 900-sample MMMU val split (TP4
-# / EP4, TRTLLM attention + MoE backends); NVFP4 (modelopt export, FP8 KV cache)
-# reuses the FP8 baseline.
-stepfun-ai/Step-3.7-Flash:
- - quant_algo: FP8_BLOCK_SCALES
- accuracy: 60
- - quant_algo: FP8_BLOCK_SCALES
- spec_dec_algo: MTP
- accuracy: 60
- - quant_algo: NVFP4
- kv_cache_quant_algo: FP8
- accuracy: 60
- - quant_algo: NVFP4
- kv_cache_quant_algo: FP8
- spec_dec_algo: MTP
- accuracy: 60
diff --git a/tests/integration/defs/accuracy/test_llm_api_pytorch.py b/tests/integration/defs/accuracy/test_llm_api_pytorch.py
index 49d7789f23e8..456f6895c7a9 100644
--- a/tests/integration/defs/accuracy/test_llm_api_pytorch.py
+++ b/tests/integration/defs/accuracy/test_llm_api_pytorch.py
@@ -8424,88 +8424,3 @@ def test_nvfp4_mtp_index_share_mtp_ar(self, tp_size, ep_size):
assert aggregate_accept_rate > 0.2, (
f"Aggregate acceptance rate {aggregate_accept_rate:.2%} "
f"below threshold 20%")
-
-
-@skip_pre_blackwell
-class TestStep3_7(LlmapiAccuracyTestHarness):
- # Step-3.7-Flash is a MoE model registered under the multimodal
- # architecture (Step3p7ForConditionalGeneration); text-only GSM8K exercises
- # the text decoder path. The custom HF config requires trust_remote_code.
- MODEL_NAME = "stepfun-ai/Step-3.7-Flash"
-
- @pytest.mark.skip_less_mpi_world_size(8)
- @pytest.mark.skip_less_device_memory(140000)
- @parametrize_with_ids("tp_size,ep_size", [(8, 8)])
- def test_auto_dtype(self, tp_size, ep_size):
- model_path = f"{llm_models_root()}/Step-3.7-Flash"
- kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.7)
- with LLM(model_path,
- tensor_parallel_size=tp_size,
- moe_expert_parallel_size=ep_size,
- kv_cache_config=kv_cache_config,
- max_seq_len=8192,
- trust_remote_code=True) as llm:
- task = GSM8K(self.MODEL_NAME)
- task.evaluate(llm)
-
- @pytest.mark.skip_less_device(4)
- @pytest.mark.skip_less_device_memory(80000)
- @parametrize_with_ids("mtp_nextn", [0, 3])
- @parametrize_with_ids("tp_size,ep_size", [(4, 4)])
- def test_fp8_block_scales(self, tp_size, ep_size, mtp_nextn):
- model_path = f"{llm_models_root()}/Step-3.7-Flash-FP8"
- kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.7,
- use_kv_cache_manager_v2=True)
- pytorch_config = dict(
- disable_overlap_scheduler=False,
- cuda_graph_config=CudaGraphConfig(enable_padding=True),
- moe_config=MoeConfig(backend="TRTLLM"),
- )
-
- mtp_config = None
- if mtp_nextn > 0:
- mtp_config = MTPDecodingConfig(max_draft_len=mtp_nextn)
-
- with LLM(model_path,
- tensor_parallel_size=tp_size,
- moe_expert_parallel_size=ep_size,
- kv_cache_config=kv_cache_config,
- max_seq_len=8192,
- attn_backend="TRTLLM",
- speculative_config=mtp_config,
- trust_remote_code=True,
- **pytorch_config) as llm:
- assert llm.args.quant_config.quant_algo == QuantAlgo.FP8_BLOCK_SCALES
- task = GSM8K(self.MODEL_NAME)
- task.evaluate(llm)
-
- @pytest.mark.skip_less_device(4)
- @pytest.mark.skip_less_device_memory(80000)
- @parametrize_with_ids("mtp_nextn", [0, 3])
- @parametrize_with_ids("tp_size,ep_size", [(4, 4)])
- def test_nvfp4(self, tp_size, ep_size, mtp_nextn):
- model_path = f"{llm_models_root()}/Step-3.7-Flash-NVFP4"
- kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.7,
- use_kv_cache_manager_v2=True)
- pytorch_config = dict(
- disable_overlap_scheduler=False,
- cuda_graph_config=CudaGraphConfig(enable_padding=True),
- moe_config=MoeConfig(backend="TRTLLM"),
- )
-
- mtp_config = None
- if mtp_nextn > 0:
- mtp_config = MTPDecodingConfig(max_draft_len=mtp_nextn)
-
- with LLM(model_path,
- tensor_parallel_size=tp_size,
- moe_expert_parallel_size=ep_size,
- kv_cache_config=kv_cache_config,
- max_seq_len=8192,
- attn_backend="TRTLLM",
- speculative_config=mtp_config,
- trust_remote_code=True,
- **pytorch_config) as llm:
- assert llm.args.quant_config.quant_algo == QuantAlgo.NVFP4
- task = GSM8K(self.MODEL_NAME)
- task.evaluate(llm)
diff --git a/tests/integration/defs/accuracy/test_llm_api_pytorch_multimodal.py b/tests/integration/defs/accuracy/test_llm_api_pytorch_multimodal.py
index 6ac701374cbd..f9c2f22880e1 100644
--- a/tests/integration/defs/accuracy/test_llm_api_pytorch_multimodal.py
+++ b/tests/integration/defs/accuracy/test_llm_api_pytorch_multimodal.py
@@ -29,7 +29,6 @@
from ..conftest import (
get_sm_version,
llm_models_root,
- parametrize_with_ids,
skip_post_blackwell_ultra,
skip_pre_blackwell,
skip_pre_hopper,
@@ -821,82 +820,3 @@ def test_auto_dtype(
sampling_params=sampling_params,
extra_evaluator_kwargs=extra_evaluator_kwargs,
)
-
-
-@skip_pre_blackwell
-class TestStep3_7(LlmapiAccuracyTestHarness):
- # Step-3.7-Flash is a reasoning VLM: a PerceptionEncoder vision tower plus a
- # MoE text decoder, registered under the Step3p7ForConditionalGeneration
- # architecture (custom HF config -> trust_remote_code). MMMU exercises the
- # vision path end to end. The model emits ... traces before
- # its answer, so we strip them and extract the final MMMU letter (same
- # handling as Kimi K2.5); preserve_caller_max_tokens keeps our larger
- # generation budget instead of lm-eval's 512-token default (too small for
- # the chain-of-thought). The text-only GSM8K path lives in
- # test_llm_api_pytorch.py::TestStep3_7.
- MODEL_NAME = "stepfun-ai/Step-3.7-Flash"
-
- # Validated with --max_input_length / --max_output_length 4096.
- sampling_params = SamplingParams(
- max_tokens=4096,
- truncate_prompt_tokens=4096,
- )
-
- EXTRA_EVALUATOR_KWARGS = dict(
- post_process_fn=strip_thinking_and_extract_mmmu_answer,
- preserve_caller_max_tokens=True,
- )
-
- kv_cache_config = KvCacheConfig(
- free_gpu_memory_fraction=0.7,
- use_kv_cache_manager_v2=True,
- )
-
- def _make_llm(self, model_path: str, mtp_nextn: int = 0):
- pytorch_config = dict(
- disable_overlap_scheduler=False,
- cuda_graph_config=CudaGraphConfig(enable_padding=False),
- moe_config=MoeConfig(backend="TRTLLM"),
- )
- mtp_config = None
- if mtp_nextn > 0:
- mtp_config = MTPDecodingConfig(max_draft_len=mtp_nextn)
- return LLM(
- model_path,
- tensor_parallel_size=4,
- moe_expert_parallel_size=4,
- kv_cache_config=self.kv_cache_config,
- max_seq_len=8192,
- attn_backend="TRTLLM",
- speculative_config=mtp_config,
- trust_remote_code=True,
- **pytorch_config,
- )
-
- @pytest.mark.skip_less_device(4)
- @pytest.mark.skip_less_device_memory(80000)
- @parametrize_with_ids("mtp_nextn", [0, 3])
- def test_fp8_block_scales(self, mtp_nextn):
- with self._make_llm(f"{llm_models_root()}/Step-3.7-Flash-FP8", mtp_nextn=mtp_nextn) as llm:
- assert llm.args.quant_config.quant_algo == QuantAlgo.FP8_BLOCK_SCALES
- task = MMMU(self.MODEL_NAME)
- task.evaluate(
- llm,
- sampling_params=self.sampling_params,
- extra_evaluator_kwargs=self.EXTRA_EVALUATOR_KWARGS,
- )
-
- @pytest.mark.skip_less_device(4)
- @pytest.mark.skip_less_device_memory(80000)
- @parametrize_with_ids("mtp_nextn", [0, 3])
- def test_nvfp4(self, mtp_nextn):
- with self._make_llm(
- f"{llm_models_root()}/Step-3.7-Flash-NVFP4", mtp_nextn=mtp_nextn
- ) as llm:
- assert llm.args.quant_config.quant_algo == QuantAlgo.NVFP4
- task = MMMU(self.MODEL_NAME)
- task.evaluate(
- llm,
- sampling_params=self.sampling_params,
- extra_evaluator_kwargs=self.EXTRA_EVALUATOR_KWARGS,
- )
diff --git a/tests/integration/test_lists/qa/llm_function_core.txt b/tests/integration/test_lists/qa/llm_function_core.txt
index 57c39a9acea2..d03a25eadf1c 100644
--- a/tests/integration/test_lists/qa/llm_function_core.txt
+++ b/tests/integration/test_lists/qa/llm_function_core.txt
@@ -772,11 +772,6 @@ accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_fp8_block_scales_early_firs
accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_w4a8_mxfp4[fp8-latency]
accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_w4a8_mxfp4[mxfp8-latency]
accuracy/test_llm_api_pytorch.py::TestSeedOss_36B::test_auto_dtype
-accuracy/test_llm_api_pytorch.py::TestStep3_7::test_auto_dtype[tp_size=8-ep_size=8] TIMEOUT (90)
-accuracy/test_llm_api_pytorch.py::TestStep3_7::test_fp8_block_scales[tp_size=4-ep_size=4-mtp_nextn=0] TIMEOUT (90)
-accuracy/test_llm_api_pytorch.py::TestStep3_7::test_fp8_block_scales[tp_size=4-ep_size=4-mtp_nextn=3] TIMEOUT (90)
-accuracy/test_llm_api_pytorch.py::TestStep3_7::test_nvfp4[tp_size=4-ep_size=4-mtp_nextn=0] TIMEOUT (90)
-accuracy/test_llm_api_pytorch.py::TestStep3_7::test_nvfp4[tp_size=4-ep_size=4-mtp_nextn=3] TIMEOUT (90)
accuracy/test_llm_api_pytorch_encode.py::TestDecoderEncode::test_decoder_encode_cuda_graph_matches_eager_logits[tinyllama-1.1b]
accuracy/test_llm_api_pytorch_encode.py::TestDecoderEncode::test_decoder_encode_matches_huggingface[gemma-3-1b]
accuracy/test_llm_api_pytorch_encode.py::TestDecoderEncode::test_decoder_encode_matches_huggingface[qwen2-7b]
@@ -806,10 +801,6 @@ accuracy/test_llm_api_pytorch_multimodal.py::TestQwen3VL_MOE::test_auto_dtype
accuracy/test_llm_api_pytorch_multimodal.py::TestQwen3_5_27B_VL::test_auto_dtype
accuracy/test_llm_api_pytorch_multimodal.py::TestQwen3_5_35B_A3B_VL::test_auto_dtype
accuracy/test_llm_api_pytorch_multimodal.py::TestQwen3_5_35B_A3B_VL::test_fp8_prequantized
-accuracy/test_llm_api_pytorch_multimodal.py::TestStep3_7::test_fp8_block_scales[mtp_nextn=0] TIMEOUT (120)
-accuracy/test_llm_api_pytorch_multimodal.py::TestStep3_7::test_fp8_block_scales[mtp_nextn=3] TIMEOUT (120)
-accuracy/test_llm_api_pytorch_multimodal.py::TestStep3_7::test_nvfp4[mtp_nextn=0] TIMEOUT (120)
-accuracy/test_llm_api_pytorch_multimodal.py::TestStep3_7::test_nvfp4[mtp_nextn=3] TIMEOUT (120)
accuracy/test_llm_api_pytorch_ray.py::TestLlama3_1_8BInstruct::test_pp2_ray
unittest/disaggregated/test_cache_transceiver_single_process.py::test_cache_transceiver_v1_masked_dsa_indexer_across_asymmetric_pp
disaggregated/test_ad_disagg.py::test_async_eagle3_full_model_handoff
diff --git a/tests/integration/test_lists/waives.txt b/tests/integration/test_lists/waives.txt
index c6451316a65e..658962bdd3a6 100644
--- a/tests/integration/test_lists/waives.txt
+++ b/tests/integration/test_lists/waives.txt
@@ -169,7 +169,6 @@ full:B300/accuracy/test_llm_api_pytorch.py::TestMiniMaxM3::test_nvfp4[use_msa=Fa
full:B300/accuracy/test_llm_api_pytorch.py::TestMistralLarge3_675B::test_nvfp4_4gpus[latency_moe_trtllm] SKIP (https://nvbugs/6529874)
full:B300/accuracy/test_llm_api_pytorch.py::TestNemotronV3Ultra::test_nvfp4_8gpus[attention_dp_off-trtllm] SKIP (https://nvbugs/6474894)
full:B300/accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_dummy_load_format SKIP (https://nvbugs/6525059)
-full:B300/accuracy/test_llm_api_pytorch.py::TestStep3_7::test_fp8_block_scales[tp_size=4-ep_size=4-mtp_nextn=3] SKIP (https://nvbugs/6539941)
full:B300/accuracy/test_llm_api_pytorch_multimodal.py::TestExaone4_5_33B::test_auto_dtype[forced_chunked_prefill] SKIP (https://nvbugs/6597570)
full:B300/accuracy/test_llm_api_pytorch_multimodal.py::TestExaone4_5_33B::test_auto_dtype[full_budget] SKIP (https://nvbugs/6597570)
full:B300/disaggregated/test_disaggregated.py::test_disaggregated_logprobs_serving[llama-3.1-8b-instruct] SKIP (https://nvbugs/6275959)