Skip to content

Add collectives configurations for different slice size #81

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
hylin2002 merged 1 commit into from
Jan 21, 2026
Merged

Add collectives configurations for different slice size #81

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 6 additions & 8 deletions Ironwood/configs/collectives/all_gather_tpu7x_2x2x1.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,11 +1,9 @@
benchmarks:
- benchmark_name: all_gather
benchmark_sweep_params:
- {matrix_dim_range: {start: 64, end: 16384, multiplier: 4}, dtype: "float32", mesh_shape: "2x2x2", ici_size_range: 8, sharding_strategy: "2x2x1", op_dimension: 1, num_runs: 5} # Parallel Replica
- {matrix_dim_range: {start: 64, end: 16384, multiplier: 4}, dtype: "float32", mesh_shape: "1x8", ici_size_range: 8, sharding_strategy: "1x8", op_dimension: 1, num_runs: 5} # Non-Parallel Replica

warmup_tries: 10
trace_dir: "../microbenchmarks/all_gather"
csv_path: "../microbenchmarks/all_gather"
xlml_metrics_dir: "../microbenchmarks/all_gather"
xla_dump_dir: "../microbenchmarks/all_gather/hlo_graphs"
- {matrix_dim_range: {start: 8, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "2x2x2", ici_size_range: 8, sharding_strategy: "2x2x1", op_dimension: 1, num_runs: 5} # Parallel Replica
- {matrix_dim_range: {start: 8, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "2x2x2", ici_size_range: 8, sharding_strategy: "2x2x2", op_dimension: 1, num_runs: 5} # Non-Parallel Replica
trace_dir: "../microbenchmarks/all_gather_tpu7x_2x2x1"
csv_path: "../microbenchmarks/all_gather_tpu7x_2x2x1"
xlml_metrics_dir: "../microbenchmarks/all_gather_tpu7x_2x2x1"
xla_dump_dir: "../microbenchmarks/all_gather_tpu7x_2x2x1/hlo_graphs"
14 changes: 6 additions & 8 deletions Ironwood/configs/collectives/all_gather_tpu7x_2x2x2.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,11 +1,9 @@
benchmarks:
- benchmark_name: all_gather
benchmark_sweep_params:
- {matrix_dim_range: {start: 64, end: 16384, multiplier: 4}, dtype: "float32", mesh_shape: "2x4x2", ici_size_range: 16, sharding_strategy: "1x4x1", op_dimension: 1, num_runs: 5} # Parallel Replica
- {matrix_dim_range: {start: 64, end: 16384, multiplier: 4}, dtype: "float32", mesh_shape: "2x2x4", ici_size_range: 16, sharding_strategy: "1x2x4", op_dimension: 1, num_runs: 5} # Non-Parallel Replica

warmup_tries: 10
trace_dir: "../microbenchmarks/all_gather"
csv_path: "../microbenchmarks/all_gather"
xlml_metrics_dir: "../microbenchmarks/all_gather"
xla_dump_dir: "../microbenchmarks/all_gather/hlo_graphs"
- {matrix_dim_range: {start: 8, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "4x2x2", ici_size_range: 16, sharding_strategy: "4x2x1", op_dimension: 1, num_runs: 5} # Parallel Replica
- {matrix_dim_range: {start: 8, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "4x2x2", ici_size_range: 16, sharding_strategy: "4x2x2", op_dimension: 1, num_runs: 5} # Non-Parallel Replica
trace_dir: "../microbenchmarks/all_gather_tpu7x_2x2x2"
csv_path: "../microbenchmarks/all_gather_tpu7x_2x2x2"
xlml_metrics_dir: "../microbenchmarks/all_gather_tpu7x_2x2x2"
xla_dump_dir: "../microbenchmarks/all_gather_tpu7x_2x2x2/hlo_graphs"
9 changes: 9 additions & 0 deletions Ironwood/configs/collectives/all_gather_tpu7x_2x2x4.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
benchmarks:
- benchmark_name: all_gather
benchmark_sweep_params:
- {matrix_dim_range: {start: 16, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "4x4x2", ici_size_range: 32, sharding_strategy: "4x4x1", op_dimension: 1, num_runs: 5} # Parallel Replica
- {matrix_dim_range: {start: 16, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "4x4x2", ici_size_range: 32, sharding_strategy: "4x4x2", op_dimension: 1, num_runs: 5} # Non-Parallel Replica
trace_dir: "../microbenchmarks/all_gather_tpu7x_2x2x4"
csv_path: "../microbenchmarks/all_gather_tpu7x_2x2x4"
xlml_metrics_dir: "../microbenchmarks/all_gather_tpu7x_2x2x4"
xla_dump_dir: "../microbenchmarks/all_gather_tpu7x_2x2x4/hlo_graphs"
9 changes: 9 additions & 0 deletions Ironwood/configs/collectives/all_gather_tpu7x_2x4x4.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
benchmarks:
- benchmark_name: all_gather
benchmark_sweep_params:
- {matrix_dim_range: {start: 32, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "8x4x2", ici_size_range: 64, sharding_strategy: "8x4x1", op_dimension: 1, num_runs: 5} # Parallel Replica
- {matrix_dim_range: {start: 32, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "8x4x2", ici_size_range: 64, sharding_strategy: "8x4x2", op_dimension: 1, num_runs: 5} # Non-Parallel Replica
trace_dir: "../microbenchmarks/all_gather_tpu7x_2x4x4"
csv_path: "../microbenchmarks/all_gather_tpu7x_2x4x4"
xlml_metrics_dir: "../microbenchmarks/all_gather_tpu7x_2x4x4"
xla_dump_dir: "../microbenchmarks/all_gather_tpu7x_2x4x4/hlo_graphs"
9 changes: 9 additions & 0 deletions Ironwood/configs/collectives/all_gather_tpu7x_4x4x4.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
benchmarks:
- benchmark_name: all_gather
benchmark_sweep_params:
- {matrix_dim_range: {start: 64, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "16x4x2", ici_size_range: 128, sharding_strategy: "16x4x1", op_dimension: 1, num_runs: 5} # Parallel Replica
- {matrix_dim_range: {start: 64, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "16x4x2", ici_size_range: 128, sharding_strategy: "16x4x2", op_dimension: 1, num_runs: 5} # Non-Parallel Replica
trace_dir: "../microbenchmarks/all_gather_tpu7x_4x4x4"
csv_path: "../microbenchmarks/all_gather_tpu7x_4x4x4"
xlml_metrics_dir: "../microbenchmarks/all_gather_tpu7x_4x4x4"
xla_dump_dir: "../microbenchmarks/all_gather_tpu7x_4x4x4/hlo_graphs"
9 changes: 9 additions & 0 deletions Ironwood/configs/collectives/all_gather_tpu7x_4x4x8.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
benchmarks:
- benchmark_name: all_gather
benchmark_sweep_params:
- {matrix_dim_range: {start: 128, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "32x4x2", ici_size_range: 256, sharding_strategy: "32x4x1", op_dimension: 1, num_runs: 5} # Parallel Replica
- {matrix_dim_range: {start: 128, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "32x4x2", ici_size_range: 256, sharding_strategy: "32x4x2", op_dimension: 1, num_runs: 5} # Non-Parallel Replica
trace_dir: "../microbenchmarks/all_gather_tpu7x_4x4x8"
csv_path: "../microbenchmarks/all_gather_tpu7x_4x4x8"
xlml_metrics_dir: "../microbenchmarks/all_gather_tpu7x_4x4x8"
xla_dump_dir: "../microbenchmarks/all_gather_tpu7x_4x4x8/hlo_graphs"
9 changes: 9 additions & 0 deletions Ironwood/configs/collectives/all_reduce_tpu7x_2x2x1.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
benchmarks:
- benchmark_name: psum
benchmark_sweep_params:
- {matrix_dim_range: {start: 8, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "2x2x2", ici_size_range: 8, sharding_strategy: "2x2x1", op_dimension: 1, num_runs: 5} # Parallel Replica
- {matrix_dim_range: {start: 8, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "2x2x2", ici_size_range: 8, sharding_strategy: "2x2x2", op_dimension: 1, num_runs: 5} # Non-Parallel Replica
trace_dir: "../microbenchmarks/psum_tpu7x_2x2x1"
csv_path: "../microbenchmarks/psum_tpu7x_2x2x1"
xlml_metrics_dir: "../microbenchmarks/psum_tpu7x_2x2x1"
xla_dump_dir: "../microbenchmarks/psum_tpu7x_2x2x1/hlo_graphs"
9 changes: 9 additions & 0 deletions Ironwood/configs/collectives/all_reduce_tpu7x_2x2x2.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
benchmarks:
- benchmark_name: psum
benchmark_sweep_params:
- {matrix_dim_range: {start: 8, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "4x2x2", ici_size_range: 16, sharding_strategy: "4x2x1", op_dimension: 1, num_runs: 5} # Parallel Replica
- {matrix_dim_range: {start: 8, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "4x2x2", ici_size_range: 16, sharding_strategy: "4x2x2", op_dimension: 1, num_runs: 5} # Non-Parallel Replica
trace_dir: "../microbenchmarks/psum_tpu7x_2x2x2"
csv_path: "../microbenchmarks/psum_tpu7x_2x2x2"
xlml_metrics_dir: "../microbenchmarks/psum_tpu7x_2x2x2"
xla_dump_dir: "../microbenchmarks/psum_tpu7x_2x2x2/hlo_graphs"
9 changes: 9 additions & 0 deletions Ironwood/configs/collectives/all_reduce_tpu7x_2x2x4.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
benchmarks:
- benchmark_name: psum
benchmark_sweep_params:
- {matrix_dim_range: {start: 16, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "4x4x2", ici_size_range: 32, sharding_strategy: "4x4x1", op_dimension: 1, num_runs: 5} # Parallel Replica
- {matrix_dim_range: {start: 16, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "4x4x2", ici_size_range: 32, sharding_strategy: "4x4x2", op_dimension: 1, num_runs: 5} # Non-Parallel Replica
trace_dir: "../microbenchmarks/psum_tpu7x_2x2x4"
csv_path: "../microbenchmarks/psum_tpu7x_2x2x4"
xlml_metrics_dir: "../microbenchmarks/psum_tpu7x_2x2x4"
xla_dump_dir: "../microbenchmarks/psum_tpu7x_2x2x4/hlo_graphs"
9 changes: 9 additions & 0 deletions Ironwood/configs/collectives/all_reduce_tpu7x_2x4x4.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
benchmarks:
- benchmark_name: psum
benchmark_sweep_params:
- {matrix_dim_range: {start: 32, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "8x4x2", ici_size_range: 64, sharding_strategy: "8x4x1", op_dimension: 1, num_runs: 5} # Parallel Replica
- {matrix_dim_range: {start: 32, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "8x4x2", ici_size_range: 64, sharding_strategy: "8x4x2", op_dimension: 1, num_runs: 5} # Non-Parallel Replica
trace_dir: "../microbenchmarks/psum_tpu7x_2x4x4"
csv_path: "../microbenchmarks/psum_tpu7x_2x4x4"
xlml_metrics_dir: "../microbenchmarks/psum_tpu7x_2x4x4"
xla_dump_dir: "../microbenchmarks/psum_tpu7x_2x4x4/hlo_graphs"
9 changes: 9 additions & 0 deletions Ironwood/configs/collectives/all_reduce_tpu7x_4x4x4.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
benchmarks:
- benchmark_name: psum
benchmark_sweep_params:
- {matrix_dim_range: {start: 64, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "16x4x2", ici_size_range: 128, sharding_strategy: "16x4x1", op_dimension: 1, num_runs: 5} # Parallel Replica
- {matrix_dim_range: {start: 64, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "16x4x2", ici_size_range: 128, sharding_strategy: "16x4x2", op_dimension: 1, num_runs: 5} # Non-Parallel Replica
trace_dir: "../microbenchmarks/psum_tpu7x_4x4x4"
csv_path: "../microbenchmarks/psum_tpu7x_4x4x4"
xlml_metrics_dir: "../microbenchmarks/psum_tpu7x_4x4x4"
xla_dump_dir: "../microbenchmarks/psum_tpu7x_4x4x4/hlo_graphs"
9 changes: 9 additions & 0 deletions Ironwood/configs/collectives/all_reduce_tpu7x_4x4x8.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
benchmarks:
- benchmark_name: psum
benchmark_sweep_params:
- {matrix_dim_range: {start: 128, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "32x4x2", ici_size_range: 256, sharding_strategy: "32x4x1", op_dimension: 1, num_runs: 5} # Parallel Replica
- {matrix_dim_range: {start: 128, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "32x4x2", ici_size_range: 256, sharding_strategy: "32x4x2", op_dimension: 1, num_runs: 5} # Non-Parallel Replica
trace_dir: "../microbenchmarks/psum_tpu7x_4x4x8"
csv_path: "../microbenchmarks/psum_tpu7x_4x4x8"
xlml_metrics_dir: "../microbenchmarks/psum_tpu7x_4x4x8"
xla_dump_dir: "../microbenchmarks/psum_tpu7x_4x4x8/hlo_graphs"
9 changes: 9 additions & 0 deletions Ironwood/configs/collectives/all_to_all_tpu7x_2x2x1.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
benchmarks:
- benchmark_name: all_to_all
benchmark_sweep_params:
- {matrix_dim_range: {start: 8, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "2x2x2", ici_size_range: 8, sharding_strategy: "2x2x1", op_dimension: 1, num_runs: 5} # Parallel Replica
- {matrix_dim_range: {start: 8, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "2x2x2", ici_size_range: 8, sharding_strategy: "2x2x2", op_dimension: 1, num_runs: 5} # Non-Parallel Replica
trace_dir: "../microbenchmarks/all_to_all_tpu7x_2x2x1"
csv_path: "../microbenchmarks/all_to_all_tpu7x_2x2x1"
xlml_metrics_dir: "../microbenchmarks/all_to_all_tpu7x_2x2x1"
xla_dump_dir: "../microbenchmarks/all_to_all_tpu7x_2x2x1/hlo_graphs"
9 changes: 9 additions & 0 deletions Ironwood/configs/collectives/all_to_all_tpu7x_2x2x2.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
benchmarks:
- benchmark_name: all_to_all
benchmark_sweep_params:
- {matrix_dim_range: {start: 8, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "4x2x2", ici_size_range: 16, sharding_strategy: "4x2x1", op_dimension: 1, num_runs: 5} # Parallel Replica
- {matrix_dim_range: {start: 8, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "4x2x2", ici_size_range: 16, sharding_strategy: "4x2x2", op_dimension: 1, num_runs: 5} # Non-Parallel Replica
trace_dir: "../microbenchmarks/all_to_all_tpu7x_2x2x2"
csv_path: "../microbenchmarks/all_to_all_tpu7x_2x2x2"
xlml_metrics_dir: "../microbenchmarks/all_to_all_tpu7x_2x2x2"
xla_dump_dir: "../microbenchmarks/all_to_all_tpu7x_2x2x2/hlo_graphs"
9 changes: 9 additions & 0 deletions Ironwood/configs/collectives/all_to_all_tpu7x_2x2x4.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
benchmarks:
- benchmark_name: all_to_all
benchmark_sweep_params:
- {matrix_dim_range: {start: 16, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "4x4x2", ici_size_range: 32, sharding_strategy: "4x4x1", op_dimension: 1, num_runs: 5} # Parallel Replica
- {matrix_dim_range: {start: 16, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "4x4x2", ici_size_range: 32, sharding_strategy: "4x4x2", op_dimension: 1, num_runs: 5} # Non-Parallel Replica
trace_dir: "../microbenchmarks/all_to_all_tpu7x_2x2x4"
csv_path: "../microbenchmarks/all_to_all_tpu7x_2x2x4"
xlml_metrics_dir: "../microbenchmarks/all_to_all_tpu7x_2x2x4"
xla_dump_dir: "../microbenchmarks/all_to_all_tpu7x_2x2x4/hlo_graphs"
9 changes: 9 additions & 0 deletions Ironwood/configs/collectives/all_to_all_tpu7x_2x4x4.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
benchmarks:
- benchmark_name: all_to_all
benchmark_sweep_params:
- {matrix_dim_range: {start: 32, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "8x4x2", ici_size_range: 64, sharding_strategy: "8x4x1", op_dimension: 1, num_runs: 5} # Parallel Replica
- {matrix_dim_range: {start: 32, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "8x4x2", ici_size_range: 64, sharding_strategy: "8x4x2", op_dimension: 1, num_runs: 5} # Non-Parallel Replica
trace_dir: "../microbenchmarks/all_to_all_tpu7x_2x4x4"
csv_path: "../microbenchmarks/all_to_all_tpu7x_2x4x4"
xlml_metrics_dir: "../microbenchmarks/all_to_all_tpu7x_2x4x4"
xla_dump_dir: "../microbenchmarks/all_to_all_tpu7x_2x4x4/hlo_graphs"
9 changes: 9 additions & 0 deletions Ironwood/configs/collectives/all_to_all_tpu7x_4x4x4.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
benchmarks:
- benchmark_name: all_to_all
benchmark_sweep_params:
- {matrix_dim_range: {start: 64, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "16x4x2", ici_size_range: 128, sharding_strategy: "16x4x1", op_dimension: 1, num_runs: 5} # Parallel Replica
- {matrix_dim_range: {start: 64, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "16x4x2", ici_size_range: 128, sharding_strategy: "16x4x2", op_dimension: 1, num_runs: 5} # Non-Parallel Replica
trace_dir: "../microbenchmarks/all_to_all_tpu7x_4x4x4"
csv_path: "../microbenchmarks/all_to_all_tpu7x_4x4x4"
xlml_metrics_dir: "../microbenchmarks/all_to_all_tpu7x_4x4x4"
xla_dump_dir: "../microbenchmarks/all_to_all_tpu7x_4x4x4/hlo_graphs"
9 changes: 9 additions & 0 deletions Ironwood/configs/collectives/all_to_all_tpu7x_4x4x8.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
benchmarks:
- benchmark_name: all_to_all
benchmark_sweep_params:
- {matrix_dim_range: {start: 128, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "32x4x2", ici_size_range: 256, sharding_strategy: "32x4x1", op_dimension: 1, num_runs: 5} # Parallel Replica
- {matrix_dim_range: {start: 128, end: 16384, multiplier: 2}, dtype: "float32", mesh_shape: "32x4x2", ici_size_range: 256, sharding_strategy: "32x4x2", op_dimension: 1, num_runs: 5} # Non-Parallel Replica
trace_dir: "../microbenchmarks/all_to_all_tpu7x_4x4x8"
csv_path: "../microbenchmarks/all_to_all_tpu7x_4x4x8"
xlml_metrics_dir: "../microbenchmarks/all_to_all_tpu7x_4x4x8"
xla_dump_dir: "../microbenchmarks/all_to_all_tpu7x_4x4x8/hlo_graphs"