Skip to content

[benchmark_inference.py] Specify tp_size to StaticCache #2784

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

Open
crcrpar wants to merge 3 commits into
base: main
Choose a base branch
from
+27 −16
Open

[benchmark_inference.py] Specify tp_size to StaticCache #2784

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
adf5377
fix sharding scheme
crcrpar Dec 3, 2025
1069a34
Merge branch 'main' into mkozuki/inference_bench_tensor_parallel_shar…
KaelanDt Jan 5, 2026
9a58218
Merge branch 'main' into mkozuki/inference_bench_tensor_parallel_shar…
crcrpar Jan 7, 2026
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
43 changes: 27 additions & 16 deletions thunder/benchmarks/benchmark_inference.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -272,50 +272,56 @@ def __init__(self, config: InferenceBenchmarkConfig):
assert all(p.device == torch.device("meta") for p in model.parameters())

tp_plan = {
"*.layers.*.self_attn.q_proj": ColwiseParallel(use_local_output=True),
"*.layers.*.self_attn.k_proj": ColwiseParallel(use_local_output=True),
"*.layers.*.self_attn.v_proj": ColwiseParallel(use_local_output=True),
"*.layers.*.self_attn.o_proj": RowwiseParallel(use_local_output=True),
"*.layers.*.feed_forward.gate_proj": ColwiseParallel(use_local_output=False),
"*.layers.*.feed_forward.up_proj": ColwiseParallel(use_local_output=False),
"*.layers.*.feed_forward.down_proj": RowwiseParallel(use_local_output=True),
"model.layers.*.self_attn.q_proj": ColwiseParallel(use_local_output=True),
"model.layers.*.self_attn.k_proj": ColwiseParallel(use_local_output=True),
"model.layers.*.self_attn.v_proj": ColwiseParallel(use_local_output=True),
"model.layers.*.self_attn.o_proj": RowwiseParallel(use_local_output=True),
"model.layers.*.feed_forward.gate_proj": ColwiseParallel(use_local_output=False),
"model.layers.*.feed_forward.up_proj": ColwiseParallel(use_local_output=False),
"model.layers.*.feed_forward.down_proj": RowwiseParallel(use_local_output=True),
}

if not self.config.disable_moe_replacement:
tp_plan.update(
{
# Custom MoE
"*.layers.*.feed_forward.shared_experts.gate_proj": ColwiseParallel(
"model.layers.*.feed_forward.shared_experts.gate_proj": ColwiseParallel(
use_local_output=False, output_layouts=Shard(2)
),
"*.layers.*.feed_forward.shared_experts.up_proj": ColwiseParallel(
"model.layers.*.feed_forward.shared_experts.up_proj": ColwiseParallel(
use_local_output=False, output_layouts=Shard(2)
),
"*.layers.*.feed_forward.shared_experts.down_proj": RowwiseParallel(),
"*.layers.*.feed_forward.routed_experts.gate_proj": GroupedLinearColwiseParallel(
"model.layers.*.feed_forward.shared_experts.down_proj": RowwiseParallel(),
"model.layers.*.feed_forward.routed_experts.gate_proj": GroupedLinearColwiseParallel(
use_local_output=False
),
"*.layers.*.feed_forward.routed_experts.up_proj": GroupedLinearColwiseParallel(
"model.layers.*.feed_forward.routed_experts.up_proj": GroupedLinearColwiseParallel(
use_local_output=False
),
"*.layers.*.feed_forward.routed_experts.down_proj": GroupedLinearRowwiseParallel(),
"model.layers.*.feed_forward.routed_experts.down_proj": GroupedLinearRowwiseParallel(),
}
)

else:
tp_plan.update(
{
# HF MoE
"*.layers.*.feed_forward.shared_expert.gate_proj": ColwiseParallel(use_local_output=False),
"*.layers.*.feed_forward.shared_expert.up_proj": ColwiseParallel(use_local_output=False),
"*.layers.*.feed_forward.shared_expert.down_proj": RowwiseParallel(use_local_output=True),
"model.layers.*.feed_forward.shared_expert.gate_proj": ColwiseParallel(use_local_output=False),
"model.layers.*.feed_forward.shared_expert.up_proj": ColwiseParallel(use_local_output=False),
"model.layers.*.feed_forward.shared_expert.down_proj": RowwiseParallel(use_local_output=True),
# TODO:Need to write ParallelStyle for HF's grouped_mm implementation.
}
)

if mesh:
model = parallelize_module(model, mesh, tp_plan)

# Sanity check - verify attention projections are sharded
assert type(model.model.layers[0].self_attn.q_proj.weight) == DTensor, "q_proj should be DTensor"
assert type(model.model.layers[0].self_attn.k_proj.weight) == DTensor, "k_proj should be DTensor"
assert type(model.model.layers[0].self_attn.v_proj.weight) == DTensor, "v_proj should be DTensor"
assert type(model.model.layers[0].self_attn.o_proj.weight) == DTensor, "o_proj should be DTensor"

# Sanity check
if not self.config.disable_moe_replacement:
assert type(model.model.layers[1].feed_forward.shared_experts.gate_proj.weight) == DTensor
Expand Down Expand Up @@ -410,12 +416,17 @@ def generate_batch(self) -> tuple[torch.Tensor, HybridChunkedCache]:
input_ids = torch.randint(0, self.vocab_size, (batch_size, input_length), device=DEVICE)
if LooseVersion(transformers.__version__) >= LooseVersion("4.55"):
# Transformers deprecated HybridChunkedCache in favour of static in 4.55.x
# NOTE: tp_size should only reflect tensor parallelism size, not total world size.
# If 2D parallelism (data parallel + tensor parallel) is added in the future,
# tp_size should be set to the tensor parallelism size only (e.g., WORLD_SIZE // DP_SIZE),
# not WORLD_SIZE, so that StaticCache correctly handles sharded KV heads.
past_key_values = StaticCache(
config=self.hf_config,
max_batch_size=input_ids.shape[0],
max_cache_len=input_ids.shape[1] + self.config.output_length,
device=DEVICE,
dtype=torch.bfloat16,
tp_size=WORLD_SIZE if mesh else 1,
)
else:
past_key_values = HybridChunkedCache(
Expand Down
Loading