[no ci] Turn off TORCHINDUCTOR_GRAPH_PARTITION

Signed-off-by: Huy Do <huydhn@gmail.com>

.ci/pytorch/test.sh

@@ -826,6 +826,24 @@ test_dynamo_benchmark() {
   local shard_id="$1"
   shift
 
+
+  ### Perf benchmark 2.9 RC4, need to reinstall detectron2 to avoid crashing when importing it
+  pip_uninstall torch torchvision torchaudio torchrec fbgemm-gpu triton pytorch-triton detectron2
+  pip_install torch==2.9.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/test/cu128
+  # Rebuild torchrec and fbgemm because they don't have RC for 2.9 yet
+  if [[ "${TEST_CONFIG}" == *torchbench* ]] && [[ "${TEST_CONFIG}" != *cpu* ]]; then
+    rm -rf dist/torchrec
+    rm -rf dist/fbgemm_gpu
+    install_torchrec_and_fbgemm
+  fi
+  # Same pinned commit as used in TorchBench
+  pip_install git+https://github.com/facebookresearch/detectron2.git@0df2d73d0013db7de629602c23cc120219b4f2b8
+  pip freeze
+
+  # Control TORCHINDUCTOR_GRAPH_PARTITION
+  export TORCHINDUCTOR_GRAPH_PARTITION=0
+
+
   if [[ "${TEST_CONFIG}" == *perf_compare* ]]; then
     test_single_dynamo_benchmark "training" "$suite" "$shard_id" --training --amp "$@"
   elif [[ "${TEST_CONFIG}" == *perf* ]]; then

.github/workflows/inductor-perf-test-nightly-h100.yml

@@ -100,11 +100,12 @@ jobs:
       cuda-arch-list: '9.0'
       test-matrix: |
         { include: [
-          { config: "inductor_huggingface_perf_cuda_h100", shard: 1, num_shards: 5, runner: "linux.aws.h100" },
-          { config: "inductor_huggingface_perf_cuda_h100", shard: 2, num_shards: 5, runner: "linux.aws.h100" },
-          { config: "inductor_huggingface_perf_cuda_h100", shard: 3, num_shards: 5, runner: "linux.aws.h100" },
-          { config: "inductor_huggingface_perf_cuda_h100", shard: 4, num_shards: 5, runner: "linux.aws.h100" },
-          { config: "inductor_huggingface_perf_cuda_h100", shard: 5, num_shards: 5, runner: "linux.aws.h100" },
+          { config: "inductor_huggingface_perf_cuda_h100", shard: 1, num_shards: 6, runner: "linux.aws.h100" },
+          { config: "inductor_huggingface_perf_cuda_h100", shard: 2, num_shards: 6, runner: "linux.aws.h100" },
+          { config: "inductor_huggingface_perf_cuda_h100", shard: 3, num_shards: 6, runner: "linux.aws.h100" },
+          { config: "inductor_huggingface_perf_cuda_h100", shard: 4, num_shards: 6, runner: "linux.aws.h100" },
+          { config: "inductor_huggingface_perf_cuda_h100", shard: 5, num_shards: 6, runner: "linux.aws.h100" },
+          { config: "inductor_huggingface_perf_cuda_h100", shard: 6, num_shards: 6, runner: "linux.aws.h100" },
           { config: "inductor_timm_perf_cuda_h100", shard: 1, num_shards: 7, runner: "linux.aws.h100" },
           { config: "inductor_timm_perf_cuda_h100", shard: 2, num_shards: 7, runner: "linux.aws.h100" },
           { config: "inductor_timm_perf_cuda_h100", shard: 3, num_shards: 7, runner: "linux.aws.h100" },

benchmarks/dynamo/common.py

@@ -3355,7 +3355,7 @@ def parse_args(args=None):
     parser.add_argument(
         "--timeout",
         type=int,
-        default=2000,
+        default=3600,
         help="timeout (second) for benchmarking.",
     )
 

benchmarks/dynamo/huggingface.py

@@ -373,6 +373,10 @@ class HuggingfaceRunner(BenchmarkRunner):
     def skip_models_due_to_control_flow(self):
         return self._skip["control_flow"]
 
+    @property
+    def skip_not_suitable_for_training_models(self):
+        return self._skip["test"]["training"]
+
     def use_larger_multiplier_for_smaller_tensor(self, name):
         return name in [
             "ElectraForQuestionAnswering",

benchmarks/dynamo/huggingface.yaml

@@ -9,10 +9,9 @@ skip:
     # Fails with even batch size = 1
     - GPTJForCausalLM
     - GPTJForQuestionAnswering
-    # Model too big
+    # Model too big or the benchmark is taking too long (timeout)
     - google/gemma-3-4b-it
     - openai/gpt-oss-20b
-    - mistralai/Mistral-7B-Instruct-v0.3
 
   device:
     cpu:
@@ -27,6 +26,16 @@ skip:
   control_flow:
     - AllenaiLongformerBase
 
+  test:
+    training:
+      - meta-llama/Llama-3.2-1B
+      - google/gemma-2-2b
+      - google/gemma-3-4b-it
+      - openai/whisper-tiny
+      - Qwen/Qwen3-0.6B
+      - mistralai/Mistral-7B-Instruct-v0.3
+      - openai/gpt-oss-20b
+
 batch_size:
   # TODO - Fails even after fake tensors
   divisors: