Skip to content

Misc: Remaining cherry-picks for 26.02.01 #2631

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
ko3n1g merged 4 commits into from
Mar 5, 2026
Merged

Misc: Remaining cherry-picks for 26.02.01 #2631

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
658522f
fix: Perf configs after refactoring (#2357)
ko3n1g Feb 15, 2026
6a41745
fix: Set flex and alltoall for perf jobs (#2393)
ko3n1g Feb 17, 2026
5c38201
qwen x86 flex backend (#2351)
malay-nagda Feb 12, 2026
0a1ebe6
DeepSeek-V3 recipes for H100 (#2312)
scsudhakaran Feb 11, 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
10 changes: 5 additions & 5 deletions scripts/performance/configs/deepseek/deepseek_llm_pretrain.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -203,18 +203,18 @@ def deepseek_v3_pretrain_config_h100(
)
precision_config = get_precision_config(precision)

pp_layout = None
if base_cfg.pp_layout:
pp_layout = base_cfg.pp_layout

cfg = pretrain_config()
cfg.mixed_precision = precision_config

# Apply model-specific settings that were previously passed as constructor args
cfg.model.pipeline_model_parallel_size = base_cfg.pipeline_model_parallel_size
cfg.model.virtual_pipeline_model_parallel_size = base_cfg.virtual_pipeline_model_parallel_size
cfg.model.moe_flex_dispatcher_backend = base_cfg.moe_flex_dispatcher_backend
cfg.model.pipeline_model_parallel_layout = pp_layout or "Et|(tt|)*30mL"
if base_cfg.pp_layout:
cfg.model.pipeline_model_parallel_layout = base_cfg.pp_layout
else:
# Recompute layout based on updated PP/VP sizes
set_deepseek_v3_pipeline_model_parallel_layout(cfg.model)

set_deepseek_v3_common_configs(cfg)
set_workload_base_configs(cfg, base_cfg)
Expand Down
8 changes: 8 additions & 0 deletions scripts/performance/configs/qwen/qwen3_llm_pretrain.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -78,6 +78,7 @@ def qwen3_235b_a22b_pretrain_config_gb300(
cfg.mixed_precision = precision_config
cfg.comm_overlap = CommOverlapConfig(tp_comm_overlap=True)
cfg.model.moe_flex_dispatcher_backend = base_cfg.moe_flex_dispatcher_backend
cfg.model.moe_token_dispatcher_type = "flex"

set_qwen3_common_configs(cfg)
set_workload_base_configs(cfg, base_cfg)
Expand All @@ -103,6 +104,7 @@ def qwen3_235b_a22b_pretrain_config_gb200(
cfg.mixed_precision = precision_config
cfg.comm_overlap = CommOverlapConfig(tp_comm_overlap=True)
cfg.model.moe_flex_dispatcher_backend = base_cfg.moe_flex_dispatcher_backend
cfg.model.moe_token_dispatcher_type = "flex"

set_qwen3_common_configs(cfg)
set_workload_base_configs(cfg, base_cfg)
Expand All @@ -128,6 +130,7 @@ def qwen3_235b_a22b_pretrain_config_b300(
cfg.mixed_precision = precision_config
cfg.comm_overlap = CommOverlapConfig(tp_comm_overlap=True)
cfg.model.moe_flex_dispatcher_backend = base_cfg.moe_flex_dispatcher_backend
cfg.model.moe_token_dispatcher_type = "alltoall"

set_qwen3_common_configs(cfg)
set_workload_base_configs(cfg, base_cfg)
Expand All @@ -153,6 +156,7 @@ def qwen3_235b_a22b_pretrain_config_b200(
cfg.mixed_precision = precision_config
cfg.comm_overlap = CommOverlapConfig(tp_comm_overlap=True)
cfg.model.moe_flex_dispatcher_backend = base_cfg.moe_flex_dispatcher_backend
cfg.model.moe_token_dispatcher_type = "alltoall"

set_qwen3_common_configs(cfg)
set_workload_base_configs(cfg, base_cfg)
Expand All @@ -178,6 +182,7 @@ def qwen3_235b_a22b_pretrain_config_h100(
cfg.mixed_precision = precision_config
cfg.comm_overlap = CommOverlapConfig(tp_comm_overlap=False)
cfg.model.moe_flex_dispatcher_backend = base_cfg.moe_flex_dispatcher_backend
cfg.model.moe_token_dispatcher_type = "alltoall"

set_qwen3_common_configs(cfg)
set_workload_base_configs(cfg, base_cfg)
Expand All @@ -203,6 +208,7 @@ def qwen3_30b_a3b_pretrain_config_gb300(
cfg.mixed_precision = precision_config
cfg.comm_overlap = CommOverlapConfig(tp_comm_overlap=True)
cfg.model.moe_flex_dispatcher_backend = base_cfg.moe_flex_dispatcher_backend
cfg.model.moe_token_dispatcher_type = "flex"

set_qwen3_common_configs(cfg)
set_workload_base_configs(cfg, base_cfg)
Expand All @@ -228,6 +234,7 @@ def qwen3_30b_a3b_pretrain_config_gb200(
cfg.mixed_precision = precision_config
cfg.comm_overlap = CommOverlapConfig(tp_comm_overlap=True)
cfg.model.moe_flex_dispatcher_backend = base_cfg.moe_flex_dispatcher_backend
cfg.model.moe_token_dispatcher_type = "flex"

set_qwen3_common_configs(cfg)
set_workload_base_configs(cfg, base_cfg)
Expand All @@ -253,6 +260,7 @@ def qwen3_30b_a3b_pretrain_config_b300(
cfg.mixed_precision = precision_config
cfg.comm_overlap = CommOverlapConfig(tp_comm_overlap=True)
cfg.model.moe_flex_dispatcher_backend = base_cfg.moe_flex_dispatcher_backend
cfg.model.moe_token_dispatcher_type = "flex"

set_qwen3_common_configs(cfg)
set_workload_base_configs(cfg, base_cfg)
Expand Down
2 changes: 2 additions & 0 deletions scripts/performance/configs/qwen/qwen3_workload_base_configs.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,13 +31,15 @@

BASE_QWEN3_235B_A22B_CONFIG = WorkloadBaseConfig(
expert_tensor_parallel_size=1,
moe_flex_dispatcher_backend="deepep",
)


BASE_QWEN3_30B_A3B_CONFIG = WorkloadBaseConfig(
expert_model_parallel_size=8,
expert_tensor_parallel_size=1,
global_batch_size=512,
moe_flex_dispatcher_backend="deepep",
)

BASE_QWEN3_NEXT_80B_A3B_CONFIG = WorkloadBaseConfig(
Expand Down
Loading