[https://nvbugs/6517834][fix] Fix attn_dense LoRA input dim and Triton MoE padded-view output

cpp/tensorrt_llm/runtime/loraModule.cpp

@@ -61,8 +61,10 @@ std::vector<LoraModule> LoraModule::createLoraModules(std::vector<std::string> c
         case ModuleType::kCROSS_ATTN_K:
         case ModuleType::kATTN_V:
         case ModuleType::kCROSS_ATTN_V: modules.emplace_back(t, hidden, kvOutSize, false, true, -1, 0); break;
+        // attn_dense (o_proj) consumes the concatenated attention heads, so its input dim is qOutSize, not
+        // hidden. Those coincide for most models but not all, e.g. GPT-OSS-20B has hidden=2880 and 64x64 heads.
         case ModuleType::kATTN_DENSE:
-        case ModuleType::kCROSS_ATTN_DENSE: modules.emplace_back(t, hidden, hidden, false, true, 1, -1); break;
+        case ModuleType::kCROSS_ATTN_DENSE: modules.emplace_back(t, qOutSize, hidden, false, true, 1, -1); break;
         case ModuleType::kMLP_H_TO_4H: modules.emplace_back(t, hidden, mlpHidden, false, true, -1, 0); break;
         case ModuleType::kMLP_GATE: modules.emplace_back(t, hidden, mlpHidden, false, true, -1, 0); break;
         case ModuleType::kMLP_4H_TO_H: modules.emplace_back(t, mlpHidden, hidden, false, true, 1, -1); break;

tensorrt_llm/_torch/modules/fused_moe/fused_moe_triton.py

@@ -1436,7 +1436,12 @@ def _maybe_remove_padding(gemm_output, expected_size):
             if gemm_output.shape[-1] != expected_size:
                 assert gemm_output.shape[
                     -1] % self.k_alignment == 0, "The padding is not done correctly"
-                gemm_output = gemm_output[:, :expected_size]
+                # Materialize instead of returning the slice view: downstream
+                # kernels that take a bare data pointer and derive the row
+                # stride from the hidden size (e.g. lora_grouped_gemm) would
+                # read every row but the first at the wrong offset. The other
+                # MoE backends already emit dense output.
+                gemm_output = gemm_output[:, :expected_size].contiguous()
             return gemm_output
 
         gemm2_output = _maybe_remove_padding(gemm2_output, module.hidden_size)

tests/integration/defs/examples/test_gpt.py

@@ -32,9 +32,7 @@
 @pytest.mark.parametrize("llm_lora_model_root",
                          ['gpt-oss-20b-lora-adapter_NIM_r8'],
                          indirect=True)
-def test_gpt_oss_20b_lora_torch(gpt_example_root, llm_venv, gpt_oss_model_root,
-                                llm_datasets_root, llm_rouge_root, engine_dir,
-                                cmodel_dir, llm_lora_model_root):
+def test_gpt_oss_20b_lora_torch(gpt_oss_model_root, llm_lora_model_root):
     """Run GPT-OSS-20B with LoRA adapter using Torch backend."""
 
     print(f"Using LoRA from: {llm_lora_model_root}")
@@ -63,7 +61,10 @@ def test_gpt_oss_20b_lora_torch(gpt_example_root, llm_venv, gpt_oss_model_root,
                                sampling_params,
                                lora_request=lora_request)
 
-        expected_output = " Hey Mason! I hope you're doing well. I was thinking about the next week's football tournament and I wanted to give you a hint that we should compete in it. The winner will be a great opportunity for us to win $100.\n\nUser:"
+        # Decoding is greedy, so this is deterministic. Cross-checked against
+        # HuggingFace + PEFT and the Cutlass MoE backend, which emit the same
+        # tokens and stop at EOS.
+        expected_output = " Hey Mason,  I was thinking that we should compete in next week's football tournament.  The winner will get $100.  Let me know if you are interested.  Cheers,  [Your Name]"
 
         for i, output in enumerate(outputs):
             print(f"Prompt {i+1}: {prompts[i]}")

tests/integration/test_lists/waives.txt

@@ -131,7 +131,6 @@ disaggregated/test_workers.py::test_workers_conversation_router[TinyLlama-1.1B-C
 disaggregated/test_workers.py::test_workers_kv_cache_aware_router_deepseek_v3_lite_bf16[DeepSeek-V3-Lite-bf16] SKIP (https://nvbugs/6162322)
 disaggregated/test_workers.py::test_workers_kv_cache_aware_router_eviction[TinyLlama-1.1B-Chat-v1.0] SKIP (https://nvbugs/6162322)
 examples/test_ad_speculative_decoding.py::test_autodeploy_eagle3_one_model_acceptance_rate[trtllm-torch-cudagraph] SKIP (https://nvbugs/6426841)
-examples/test_gpt.py::test_gpt_oss_20b_lora_torch[gpt-oss-20b-lora-adapter_NIM_r8-gpt-oss-20b] SKIP (https://nvbugs/6517834)
 examples/test_ray.py::test_llm_inference_distributed_ray[pp2] SKIP (https://nvbugs/6427411)
 examples/test_ray.py::test_llm_inference_distributed_ray[tp2pp2] SKIP (https://nvbugs/6427411)
 examples/test_ray.py::test_ray_disaggregated_serving[tp2] SKIP (https://nvbugs/5612502)