[https://nvbugs/5744427][fix] Make Gemma3 multimodal test fp8

Signed-off-by: Balaram Buddharaju <[email protected]>

Author: brb-nv

tests/integration/defs/accuracy/references/mmmu.yaml

@@ -1,5 +1,8 @@
 google/gemma-3-27b-it:
   - accuracy: 52.0
+  - quant_algo: FP8
+    kv_cache_quant_algo: FP8
+    accuracy: 50.0
 Qwen/Qwen2-VL-7B-Instruct:
   - accuracy: 48.44
 Qwen/Qwen2.5-VL-7B-Instruct:

tests/integration/defs/accuracy/test_llm_api_pytorch_multimodal.py

@@ -220,7 +220,8 @@ def test_auto_dtype(self):
 @skip_post_blackwell
 class TestGemma3_27BInstruct(LlmapiAccuracyTestHarness):
     MODEL_NAME = "google/gemma-3-27b-it"
-    MODEL_PATH = f"{llm_models_root()}/gemma/gemma-3-27b-it/"
+    # Note: This has only the LLM part quantized. Vision part is in bfloat16.
+    MODEL_PATH = f"{llm_models_root()}/gemma/gemma-3-27b-it-fp8/"
     MAX_NUM_TOKENS = 12800
 
     sampling_params = SamplingParams(
@@ -232,9 +233,10 @@ class TestGemma3_27BInstruct(LlmapiAccuracyTestHarness):
         enable_block_reuse=False,
         enable_partial_reuse=False,
         free_gpu_memory_fraction=0.4,
+        dtype="fp8",
     )
 
-    def test_auto_dtype(self):
+    def test_fp8_prequantized(self):
         # Gemma3 VLM needs FlashInfer attention backend for custom mask support.
         with LLM(
             self.MODEL_PATH,

tests/integration/test_lists/qa/llm_function_core.txt

@@ -675,7 +675,7 @@ accuracy/test_llm_api_pytorch_multimodal.py::TestNVILA_8B::test_auto_dtype
 accuracy/test_llm_api_pytorch_multimodal.py::TestVILA1_5_3B::test_auto_dtype
 accuracy/test_llm_api_pytorch_multimodal.py::TestNemotron_Nano_12B_V2_VL::test_auto_dtype
 accuracy/test_llm_api_pytorch_multimodal.py::TestPhi4MMFusedVisionLora::test_auto_dtype
-accuracy/test_llm_api_pytorch_multimodal.py::TestGemma3_27BInstruct::test_auto_dtype
+accuracy/test_llm_api_pytorch_multimodal.py::TestGemma3_27BInstruct::test_fp8_prequantized
 accuracy/test_llm_api_pytorch_multimodal.py::TestQwen3VL_MOE::test_auto_dtype
 
 test_e2e.py::test_llama_e2e[use_cpp_session-remove_input_padding-]

tests/integration/test_lists/qa/llm_function_core_sanity.txt

@@ -236,7 +236,7 @@ accuracy/test_llm_api_pytorch_multimodal.py::TestQwen2_5_VL_7B::test_auto_dtype
 accuracy/test_llm_api_pytorch_multimodal.py::TestLlava_V1_6_Mistral_7B::test_auto_dtype
 accuracy/test_llm_api_pytorch_multimodal.py::TestNVILA_8B::test_auto_dtype
 accuracy/test_llm_api_pytorch_multimodal.py::TestVILA1_5_3B::test_auto_dtype
-accuracy/test_llm_api_pytorch_multimodal.py::TestGemma3_27BInstruct::test_auto_dtype
+accuracy/test_llm_api_pytorch_multimodal.py::TestGemma3_27BInstruct::test_fp8_prequantized
 
 disaggregated/test_disaggregated.py::test_disaggregated_cache_aware_balance[TinyLlama-1.1B-Chat-v1.0]
 disaggregated/test_disaggregated.py::test_disaggregated_cache_aware_balance[TinyLlama-1.1B-Chat-v1.0]

tests/integration/test_lists/qa/llm_function_nim.txt

@@ -241,7 +241,7 @@ accuracy/test_llm_api_pytorch.py::TestStarcoder2_15B::test_auto_dtype
 accuracy/test_llm_api_pytorch_multimodal.py::TestQwen2_VL_7B::test_auto_dtype
 accuracy/test_llm_api_pytorch_multimodal.py::TestQwen2_5_VL_7B::test_auto_dtype
 accuracy/test_llm_api_pytorch_multimodal.py::TestLlava_V1_6_Mistral_7B::test_auto_dtype
-accuracy/test_llm_api_pytorch_multimodal.py::TestGemma3_27BInstruct::test_auto_dtype
+accuracy/test_llm_api_pytorch_multimodal.py::TestGemma3_27BInstruct::test_fp8_prequantized
 accuracy/test_disaggregated_serving.py::TestGPTOSS::test_auto_dtype[True]
 accuracy/test_disaggregated_serving.py::TestGPTOSS::test_auto_dtype[False]
 

tests/integration/test_lists/test-db/l0_h100.yml

@@ -71,6 +71,7 @@ l0_h100:
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_chunked_prefill[quant_dtype=none-kv_cache_reuse=False-fp8kv=False-overlap_scheduler=True]
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_chunked_prefill[quant_dtype=fp8-kv_cache_reuse=True-fp8kv=True-overlap_scheduler=True]
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_dummy_load_format
+  - accuracy/test_llm_api_pytorch_multimodal.py::TestGemma3_27BInstruct::test_fp8_prequantized
   - accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_fp8_block_scales[latency]
   - accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_dummy_load_format
   - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_fp8[latency-torch_compile=False]
@@ -264,7 +265,6 @@ l0_h100:
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_fp8_block_scales[mtp=vanilla-fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False]
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_no_kv_cache_reuse[quant_dtype=none-mtp_nextn=2-fp8kv=False-attention_dp=True-cuda_graph=True-overlap_scheduler=True]
   - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_guided_decoding[xgrammar-mtp_nextn=0]
-  - accuracy/test_llm_api_pytorch_multimodal.py::TestGemma3_27BInstruct::test_auto_dtype
   - accuracy/test_llm_api_pytorch.py::TestGemma3_1BInstruct::test_fp8_prequantized
   - accuracy/test_llm_api_pytorch.py::TestGemma3_27BInstruct::test_fp8_prequantized
   - accuracy/test_llm_api_pytorch.py::TestGemma3_27BInstruct::test_auto_dtype TIMEOUT (90)