Skip to content

[TRTLLM-15102][fix] Prune VILA family models #17675

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
xinhe-nv merged 1 commit into from
Aug 14, 2026
Merged

[TRTLLM-15102][fix] Prune VILA family models #17675

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 0 additions & 4 deletions tests/integration/defs/accuracy/references/mmmu.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,10 +42,6 @@ Qwen/Qwen2.5-VL-7B-Instruct:
accuracy: 40.67
nvidia/Nano-v2-VLM:
- accuracy: 43.78
Efficient-Large-Model/NVILA-8B:
- accuracy: 47.77
Efficient-Large-Model/VILA1.5-3b:
- accuracy: 32.33
# MMMU for Nemotron-Nano-* models require reasoning on.
# While enabling reasoning for current test harness is not supported, the metric here is for model
# regression testing, not an indicator of official benchmark results.
Expand Down
57 changes: 0 additions & 57 deletions tests/integration/defs/accuracy/test_llm_api_pytorch_multimodal.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -143,63 +143,6 @@ def test_auto_dtype(self, enable_chunked_prefill, max_num_tokens):
task.evaluate(llm, sampling_params=self.sampling_params)


@skip_pre_hopper
class TestNVILA_8B(LlmapiAccuracyTestHarness):
MODEL_NAME = "Efficient-Large-Model/NVILA-8B"
MODEL_PATH = f"{llm_models_root()}/vila/NVILA-8B"
MAX_NUM_TOKENS = 16384

# NOTE: MMMU adds <|endoftext|> to the stop token.
sampling_params = SamplingParams(
max_tokens=MMMU.MAX_OUTPUT_LEN,
truncate_prompt_tokens=MMMU.MAX_INPUT_LEN,
stop="<|endoftext|>",
)

kv_cache_config = KvCacheConfig(
free_gpu_memory_fraction=0.6,
# NOTE: VILA models do not support block reuse.
enable_block_reuse=False,
)

def test_auto_dtype(self):
with LLM(
self.MODEL_PATH,
max_num_tokens=self.MAX_NUM_TOKENS,
kv_cache_config=self.kv_cache_config,
) as llm:
task = MMMU(self.MODEL_NAME)
task.evaluate(llm, sampling_params=self.sampling_params)


class TestVILA1_5_3B(LlmapiAccuracyTestHarness):
MODEL_NAME = "Efficient-Large-Model/VILA1.5-3b"
MODEL_PATH = f"{llm_models_root()}/vila/VILA1.5-3b"
MAX_NUM_TOKENS = 16384

# NOTE: MMMU adds <|endoftext|> to the stop token.
sampling_params = SamplingParams(
max_tokens=MMMU.MAX_OUTPUT_LEN,
truncate_prompt_tokens=MMMU.MAX_INPUT_LEN,
stop="<|endoftext|>",
)

kv_cache_config = KvCacheConfig(
free_gpu_memory_fraction=0.6,
# NOTE: VILA models do not support block reuse.
enable_block_reuse=False,
)

def test_auto_dtype(self):
with LLM(
self.MODEL_PATH,
max_num_tokens=self.MAX_NUM_TOKENS,
kv_cache_config=self.kv_cache_config,
) as llm:
task = MMMU(self.MODEL_NAME)
task.evaluate(llm, sampling_params=self.sampling_params)


@skip_pre_hopper
class TestNemotron_Nano_12B_V2_VL(LlmapiAccuracyTestHarness):
MODEL_NAME = "nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16"
Expand Down
2 changes: 0 additions & 2 deletions tests/integration/defs/conftest.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -560,8 +560,6 @@ def multimodal_model_root(request, llm_venv):
assert models_root, "Did you set LLM_MODELS_ROOT?"

tllm_model_name = request.param
if "VILA" in tllm_model_name:
models_root = os.path.join(llm_models_root(), "vila")
if "cogvlm-chat" in tllm_model_name:
models_root = os.path.join(llm_models_root(), "cogvlm-chat")
if "video-neva" in tllm_model_name:
Expand Down
2 changes: 0 additions & 2 deletions tests/integration/test_lists/qa/llm_function_core.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -794,7 +794,6 @@ accuracy/test_llm_api_pytorch_multimodal.py::TestGemma3_27BInstruct::test_fp8_pr
accuracy/test_llm_api_pytorch_multimodal.py::TestKimiK25::test_nvfp4[dep8]
accuracy/test_llm_api_pytorch_multimodal.py::TestMistralLarge3_675B::test_nvfp4_4gpus[latency_moe_trtllm]
accuracy/test_llm_api_pytorch_multimodal.py::TestMistralSmall24B::test_auto_dtype[forced_chunked_prefill]
accuracy/test_llm_api_pytorch_multimodal.py::TestNVILA_8B::test_auto_dtype
accuracy/test_llm_api_pytorch_multimodal.py::TestNanoV3Omni::test_auto_dtype[bf16]
accuracy/test_llm_api_pytorch_multimodal.py::TestNanoV3Omni::test_auto_dtype[fp8]
accuracy/test_llm_api_pytorch_multimodal.py::TestNanoV3Omni::test_auto_dtype[fp8_mmmu_encoder_cuda_graph]
Expand All @@ -811,7 +810,6 @@ accuracy/test_llm_api_pytorch_multimodal.py::TestStep3_7::test_fp8_block_scales[
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_multimodal.py::TestVILA1_5_3B::test_auto_dtype
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
Expand Down
1 change: 0 additions & 1 deletion tests/integration/test_lists/qa/llm_function_rtx6k.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -177,4 +177,3 @@ accuracy/test_llm_api_pytorch.py::TestQwen3_5_4B::test_bf16
accuracy/test_llm_api_pytorch.py::TestQwen3_5_4B::test_dflash
accuracy/test_llm_api_pytorch.py::TestQwen3_5_4B::test_fp8
accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_dflash
accuracy/test_llm_api_pytorch_multimodal.py::TestNVILA_8B::test_auto_dtype
1 change: 0 additions & 1 deletion tests/integration/test_lists/test-db/l0_l40s.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,6 @@ l0_l40s:
# ------------- PyTorch tests ---------------
# Multimodal modeling tests
- unittest/_torch/modeling -k "modeling_siglip"
- unittest/_torch/modeling -k "modeling_vila"
- unittest/_torch/modeling -k "modeling_nemotron_nano_v2_vl"
- unittest/_torch/modeling/test_modeling_qwen2_5vl.py::TestQwen2_5_VL::test_all
- unittest/_torch/modeling/test_modeling_qwen3vl_moe.py::TestQwen3VLMoe::test_all
Expand Down
2 changes: 0 additions & 2 deletions tests/integration/test_lists/waives.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -200,8 +200,6 @@ full:GB300/accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_fp8_moe_d
full:GB300/accuracy/test_llm_api_pytorch_multimodal.py::TestExaone4_5_33B::test_auto_dtype[forced_chunked_prefill] SKIP (https://nvbugs/6597570)
full:GB300/accuracy/test_llm_api_pytorch_multimodal.py::TestExaone4_5_33B::test_auto_dtype[full_budget] SKIP (https://nvbugs/6597570)
full:GB300/accuracy/test_llm_api_pytorch_multimodal.py::TestGemma3_27BInstruct::test_fp8_prequantized SKIP (https://nvbugs/6479708)
full:GB300/accuracy/test_llm_api_pytorch_multimodal.py::TestNVILA_8B::test_auto_dtype SKIP (https://nvbugs/6597570)
full:GB300/accuracy/test_llm_api_pytorch_multimodal.py::TestVILA1_5_3B::test_auto_dtype SKIP (https://nvbugs/6597570)
full:GB300/disaggregated/test_disaggregated.py::test_disaggregated_deepseek_v3_lite_fp8_attention_dp_overlap[DeepSeek-V3-Lite-fp8] SKIP (https://nvbugs/6581064)
full:GB300/perf/test_perf_sanity.py::test_e2e[disagg_upload-gen_only-gb300_kimi-k25-thinking-fp4_8k1k_con4_ctx1_dep4_gen1_tep8_eplb0_mtp3_ccb-NIXL] SKIP (https://nvbugs/6572464)
full:H100/accuracy/test_disaggregated_serving.py::TestDeepSeekV3Lite::test_auto_dtype[mtp_nextn=2-overlap_scheduler=False] SKIP (https://nvbugs/6313072)
Expand Down
7 changes: 0 additions & 7 deletions tests/unittest/_torch/attention/model_attn_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -248,13 +248,6 @@ class ModelAttnConfig:
num_kv_heads=32,
head_dim=128,
),
ModelAttnConfig(
"vila1_5_3b_mha",
"VILA1.5-3B LLM",
num_heads=20,
num_kv_heads=20,
head_dim=128,
),
ModelAttnConfig(
"llama2_13b_qwen1_5_14b_mha",
"Llama-2-13B, Qwen1.5-14B",
Expand Down
Loading
Loading