Skip to content

Updated amax_sync test to set const weights based on rank #451

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
merged 2 commits into from
Oct 21, 2025
Merged

Updated amax_sync test to set const weights based on rank #451

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
bed2ecd
updated test to set const weights to generate diff amax
kinjalpatel27 Oct 17, 2025
d7dd6ad
updated Changelog for MoE support
kinjalpatel27 Oct 21, 2025
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
17 changes: 16 additions & 1 deletion tests/gpu/torch/quantization/plugins/test_megatron.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -673,7 +673,22 @@ def _test_expert_model_parallel_amax_sync(
num_moe_experts=8,
transformer_impl="modelopt",
)
prompt_tokens = torch.randint(0, model.vocab_size, (2, model.max_sequence_length)).cuda()

# Initialize ALL weights based on rank to produce different amax values
# to produce different amax values across ranks that need synchronization
weight_idx = 0
for name, param in model.named_parameters():
# Skip embeddings and any parameters without 'weight' in the name
if "embedding" in name.lower() or "weight" not in name.lower():
continue

if param.requires_grad and param.dim() >= 2: # Only weight matrices, not biases
# Different constant value based on rank and parameter index
const_val = 0.1 + (rank * 0.5) + (weight_idx * 0.05)
param.data.fill_(const_val)
weight_idx += 1

prompt_tokens = (torch.ones((2, model.max_sequence_length)) * 0.05 + rank * 0.5).cuda().long()

# force all expert routing
for module in model.modules():
Expand Down