[None][feat] Add support for DeepSeek v3.2 tests

tests/integration/defs/perf/disagg/test_configs/wideep/perf/deepseek-v32-fp4_1k1k_ctx1_gen1_dep32_bs32_eplb288_mtp0_ccb-NIXL.yaml

@@ -43,13 +43,12 @@ environment:
 profiling:
   nsys_on: false
 accuracy:
-  enable_accuracy_test: false
+  enable_accuracy_test: false  # Set to true to enable accuracy evaluation
   model: local-completions
   tasks: gsm8k
   model_args_extra: num_concurrent=512,max_retries=3,tokenized_requests=false,timeout=1200,max_gen_toks=256,max_length=4096
 worker_config:
   gen:
-    enable_layerwise_nvtx_marker: true
     tensor_parallel_size: 32
     moe_expert_parallel_size: 32
     enable_attention_dp: true
@@ -80,17 +79,20 @@ worker_config:
       free_gpu_memory_fraction: 0.9
       dtype: fp8
     moe_config:
-      backend: WIDEEP
-      load_balancer:
-        num_slots: 288
-        layer_updates_per_iter: 1
+      backend: CUTEDSL
+      use_low_precision_moe_combine: true
+    nvfp4_gemm_config:
+      allowed_backends:
+      - cutlass
+      - cublaslt
+      - cutedsl
+      - cuda_core
     cache_transceiver_config:
       max_tokens_in_buffer: 4608
-      backend: NIXL
+      backend: NIXLf
     stream_interval: 20
     num_postprocess_workers: 4
   ctx:
-    enable_layerwise_nvtx_marker: true
     max_batch_size: 4
     max_num_tokens: 4608
     max_seq_len: 2251
@@ -101,6 +103,8 @@ worker_config:
     print_iter_log: true
     cuda_graph_config: null
     disable_overlap_scheduler: true
+    moe_config:
+      backend: TRTLLM
     kv_cache_config:
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.85

tests/integration/defs/perf/disagg/test_configs/wideep/perf/deepseek-v32-fp4_1k1k_ctx2_gen1_dep16_bs128_eplb288_mtp3_ccb-NIXL.yaml

@@ -49,7 +49,6 @@ accuracy:
   model_args_extra: num_concurrent=512,max_retries=3,tokenized_requests=false,timeout=1200,max_gen_toks=256,max_length=4096
 worker_config:
   gen:
-    enable_layerwise_nvtx_marker: true
     tensor_parallel_size: 16
     moe_expert_parallel_size: 16
     enable_attention_dp: true
@@ -80,10 +79,14 @@ worker_config:
       free_gpu_memory_fraction: 0.9
       dtype: fp8
     moe_config:
-      backend: WIDEEP
-      load_balancer:
-        num_slots: 288
-        layer_updates_per_iter: 1
+      backend: CUTEDSL
+      use_low_precision_moe_combine: true
+    nvfp4_gemm_config:
+      allowed_backends:
+      - cutlass
+      - cublaslt
+      - cutedsl
+      - cuda_core
     cache_transceiver_config:
       max_tokens_in_buffer: 4608
       backend: NIXL
@@ -93,7 +96,6 @@ worker_config:
       decoding_type: MTP
       num_nextn_predict_layers: 3
   ctx:
-    enable_layerwise_nvtx_marker: true
     max_batch_size: 4
     max_num_tokens: 4608
     max_seq_len: 2251
@@ -104,6 +106,8 @@ worker_config:
     print_iter_log: true
     cuda_graph_config: null
     disable_overlap_scheduler: true
+    moe_config:
+      backend: TRTLLM
     kv_cache_config:
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.85

tests/integration/defs/perf/disagg/test_configs/wideep/perf/deepseek-v32-fp4_1k1k_ctx2_gen1_dep48_bs16_eplb288_mtp3_ccb-DEFAULT.yaml

@@ -50,7 +50,6 @@ accuracy:
   model_args_extra: num_concurrent=512,max_retries=3,tokenized_requests=false,timeout=1200,max_gen_toks=256,max_length=4096
 worker_config:
   gen:
-    enable_layerwise_nvtx_marker: true
     tensor_parallel_size: 48
     moe_expert_parallel_size: 48
     enable_attention_dp: true
@@ -81,16 +80,19 @@ worker_config:
       free_gpu_memory_fraction: 0.7
       dtype: fp8
     moe_config:
-      backend: WIDEEP
-      load_balancer:
-        num_slots: 288
-        layer_updates_per_iter: 1
+      backend: CUTEDSL
+      use_low_precision_moe_combine: true
+    nvfp4_gemm_config:
+      allowed_backends:
+      - cutlass
+      - cublaslt
+      - cutedsl
+      - cuda_core
     cache_transceiver_config:
       max_tokens_in_buffer: 8320
       backend: DEFAULT
     stream_interval: 20
   ctx:
-    enable_layerwise_nvtx_marker: true
     max_batch_size: 4
     max_num_tokens: 4480
     max_seq_len: 2176
@@ -101,6 +103,8 @@ worker_config:
     print_iter_log: true
     cuda_graph_config: null
     disable_overlap_scheduler: true
+    moe_config:
+      backend: TRTLLM
     kv_cache_config:
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.85

tests/integration/defs/perf/disagg/test_configs/wideep/perf/deepseek-v32-fp4_8k1k_ctx2_gen1_dep32_bs128_eplb288_mtp3_ccb-DEFAULT.yaml

@@ -49,7 +49,6 @@ accuracy:
   model_args_extra: num_concurrent=512,max_retries=3,tokenized_requests=false,timeout=1200,max_gen_toks=256,max_length=4096
 worker_config:
   gen:
-    enable_layerwise_nvtx_marker: true
     tensor_parallel_size: 32
     moe_expert_parallel_size: 32
     enable_attention_dp: true
@@ -81,10 +80,14 @@ worker_config:
       free_gpu_memory_fraction: 0.6
       dtype: fp8
     moe_config:
-      backend: WIDEEP
-      load_balancer:
-        num_slots: 288
-        layer_updates_per_iter: 1
+      backend: CUTEDSL
+      use_low_precision_moe_combine: true
+    nvfp4_gemm_config:
+      allowed_backends:
+      - cutlass
+      - cublaslt
+      - cutedsl
+      - cuda_core
     cache_transceiver_config:
       max_tokens_in_buffer: 8448
       backend: DEFAULT
@@ -94,7 +97,6 @@ worker_config:
       decoding_type: MTP
       num_nextn_predict_layers: 3
   ctx:
-    enable_layerwise_nvtx_marker: true
     max_batch_size: 1
     max_num_tokens: 8448
     max_seq_len: 9423
@@ -109,6 +111,8 @@ worker_config:
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.75
       dtype: fp8
+    moe_config:
+      backend: TRTLLM
     cache_transceiver_config:
       max_tokens_in_buffer: 8448
       backend: DEFAULT

tests/integration/defs/perf/disagg/test_configs/wideep/perf/deepseek-v32-fp4_8k1k_ctx6_gen1_dep16_bs64_eplb288_mtp0_ccb-NIXL.yaml

@@ -49,7 +49,6 @@ accuracy:
   model_args_extra: num_concurrent=512,max_retries=3,tokenized_requests=false,timeout=1200,max_gen_toks=256,max_length=4096
 worker_config:
   gen:
-    enable_layerwise_nvtx_marker: true
     tensor_parallel_size: 16
     moe_expert_parallel_size: 16
     enable_attention_dp: true
@@ -80,17 +79,20 @@ worker_config:
       free_gpu_memory_fraction: 0.7
       dtype: fp8
     moe_config:
-      backend: WIDEEP
-      load_balancer:
-        num_slots: 288
-        layer_updates_per_iter: 1
+      backend: CUTEDSL
+      use_low_precision_moe_combine: true
+    nvfp4_gemm_config:
+      allowed_backends:
+      - cutlass
+      - cublaslt
+      - cutedsl
+      - cuda_core
     cache_transceiver_config:
       max_tokens_in_buffer: 8448
       backend: NIXL
     stream_interval: 20
     num_postprocess_workers: 4
   ctx:
-    enable_layerwise_nvtx_marker: true
     max_batch_size: 1
     max_num_tokens: 8448
     max_seq_len: 9419
@@ -105,6 +107,8 @@ worker_config:
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.75
       dtype: fp8
+    moe_config:
+      backend: TRTLLM
     cache_transceiver_config:
       max_tokens_in_buffer: 8448
       backend: NIXL

tests/integration/defs/perf/disagg/test_configs/wideep/perf/deepseek-v32-fp4_8k1k_ctx8_gen1_dep32_bs16_eplb288_mtp3_ccb-NIXL.yaml

@@ -49,7 +49,6 @@ accuracy:
   model_args_extra: num_concurrent=512,max_retries=3,tokenized_requests=false,timeout=1200,max_gen_toks=256,max_length=4096
 worker_config:
   gen:
-    enable_layerwise_nvtx_marker: true
     tensor_parallel_size: 32
     moe_expert_parallel_size: 32
     enable_attention_dp: true
@@ -80,10 +79,14 @@ worker_config:
       free_gpu_memory_fraction: 0.7
       dtype: fp8
     moe_config:
-      backend: WIDEEP
-      load_balancer:
-        num_slots: 288
-        layer_updates_per_iter: 1
+      backend: CUTEDSL
+      use_low_precision_moe_combine: true
+    nvfp4_gemm_config:
+      allowed_backends:
+      - cutlass
+      - cublaslt
+      - cutedsl
+      - cuda_core
     cache_transceiver_config:
       max_tokens_in_buffer: 8448
       backend: NIXL
@@ -93,7 +96,6 @@ worker_config:
       decoding_type: MTP
       num_nextn_predict_layers: 3
   ctx:
-    enable_layerwise_nvtx_marker: true
     max_batch_size: 1
     max_num_tokens: 8448
     max_seq_len: 9419
@@ -108,6 +110,8 @@ worker_config:
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.75
       dtype: fp8
+    moe_config:
+      backend: TRTLLM
     cache_transceiver_config:
       max_tokens_in_buffer: 8448
       backend: NIXL