Increase the number of perf benchmark shards (pytorch#145534)

Per the discussion on pytorch#140332 (comment), this adds 2 more shards for HF, 2 more for TorchBench, and 1 more for TIMM.

Pull Request resolved: pytorch#145534
Approved by: https://github.com/jeanschmidt

Author: huydhn

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

@@ -88,18 +88,23 @@ jobs:
       cuda-arch-list: '8.0'
       test-matrix: |
         { include: [
-          { config: "inductor_huggingface_perf", shard: 1, num_shards: 3, runner: "linux.aws.a100" },
-          { config: "inductor_huggingface_perf", shard: 2, num_shards: 3, runner: "linux.aws.a100" },
-          { config: "inductor_huggingface_perf", shard: 3, num_shards: 3, runner: "linux.aws.a100" },
-          { config: "inductor_timm_perf", shard: 1, num_shards: 5, runner: "linux.aws.a100" },
-          { config: "inductor_timm_perf", shard: 2, num_shards: 5, runner: "linux.aws.a100" },
-          { config: "inductor_timm_perf", shard: 3, num_shards: 5, runner: "linux.aws.a100" },
-          { config: "inductor_timm_perf", shard: 4, num_shards: 5, runner: "linux.aws.a100" },
-          { config: "inductor_timm_perf", shard: 5, num_shards: 5, runner: "linux.aws.a100" },
-          { config: "inductor_torchbench_perf", shard: 1, num_shards: 4, runner: "linux.aws.a100" },
-          { config: "inductor_torchbench_perf", shard: 2, num_shards: 4, runner: "linux.aws.a100" },
-          { config: "inductor_torchbench_perf", shard: 3, num_shards: 4, runner: "linux.aws.a100" },
-          { config: "inductor_torchbench_perf", shard: 4, num_shards: 4, runner: "linux.aws.a100" },
+          { config: "inductor_huggingface_perf", shard: 1, num_shards: 5, runner: "linux.aws.a100" },
+          { config: "inductor_huggingface_perf", shard: 2, num_shards: 5, runner: "linux.aws.a100" },
+          { config: "inductor_huggingface_perf", shard: 3, num_shards: 5, runner: "linux.aws.a100" },
+          { config: "inductor_huggingface_perf", shard: 4, num_shards: 5, runner: "linux.aws.a100" },
+          { config: "inductor_huggingface_perf", shard: 5, num_shards: 5, runner: "linux.aws.a100" },
+          { config: "inductor_timm_perf", shard: 1, num_shards: 6, runner: "linux.aws.a100" },
+          { config: "inductor_timm_perf", shard: 2, num_shards: 6, runner: "linux.aws.a100" },
+          { config: "inductor_timm_perf", shard: 3, num_shards: 6, runner: "linux.aws.a100" },
+          { config: "inductor_timm_perf", shard: 4, num_shards: 6, runner: "linux.aws.a100" },
+          { config: "inductor_timm_perf", shard: 5, num_shards: 6, runner: "linux.aws.a100" },
+          { config: "inductor_timm_perf", shard: 6, num_shards: 6, runner: "linux.aws.a100" },
+          { config: "inductor_torchbench_perf", shard: 1, num_shards: 6, runner: "linux.aws.a100" },
+          { config: "inductor_torchbench_perf", shard: 2, num_shards: 6, runner: "linux.aws.a100" },
+          { config: "inductor_torchbench_perf", shard: 3, num_shards: 6, runner: "linux.aws.a100" },
+          { config: "inductor_torchbench_perf", shard: 4, num_shards: 6, runner: "linux.aws.a100" },
+          { config: "inductor_torchbench_perf", shard: 5, num_shards: 6, runner: "linux.aws.a100" },
+          { config: "inductor_torchbench_perf", shard: 6, num_shards: 6, runner: "linux.aws.a100" },
         ]}
       selected-test-configs: ${{ inputs.benchmark_configs }}
     secrets: inherit