Skip to content

Unify policy-loss metrics config and add GSPO metrics #555

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
jlamypoirier merged 4 commits into from
Jul 8, 2026
+434 −141
Merged

Unify policy-loss metrics config and add GSPO metrics #555

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
7b8ca77
Unify policy-loss metrics config and add GSPO metrics
jlamypoirier Jul 7, 2026
e4973fb
GSPO metrics: log num_tokens, dedupe loss defs, add distributed test
jlamypoirier Jul 8, 2026
fed92da
Dedupe GRPO/GSPO metrics into a shared core
jlamypoirier Jul 8, 2026
1c2005f
Fine-review fixes: naming and annotations
jlamypoirier Jul 8, 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
22 changes: 11 additions & 11 deletions fast_llm/layers/language_model/loss/config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -205,7 +205,7 @@ def loss_class(self) -> "type[LanguageModelZLoss]":
return LanguageModelZLoss


class GRPOMetricsLevel(enum.StrEnum):
class PolicyMetricsLevel(enum.StrEnum):
none = "none"
basic = "basic"
with_entropy = "with_entropy"
Expand All @@ -219,6 +219,16 @@ class LanguageModelPolicyGradientLossConfig(LanguageModelLossConfig):

epsilon_low: float = Field(default=0.2, desc="Lower clip parameter for ratio of log probs")
epsilon_high: float = Field(default=0.2, desc="Upper clip parameter for ratio of log probs")
metrics: PolicyMetricsLevel = Field(
default=PolicyMetricsLevel.none,
desc=(
"Additional diagnostic metrics to log. "
"`basic`: importance-ratio, KL and advantage statistics. "
"`with_entropy`: also log the policy entropy. "
"Not supported with pipeline_parallel > 1."
),
hint=FieldHint.feature,
)

@property
def loss_class(self) -> "type[LanguageModelPolicyGradientLoss]":
Expand All @@ -236,16 +246,6 @@ class LanguageModelGRPOLossConfig(LanguageModelPolicyGradientLossConfig):
desc="Enable triton implementation. Default: use if available.",
hint=FieldHint.expert,
)
metrics: GRPOMetricsLevel = Field(
default=GRPOMetricsLevel.none,
desc=(
"Additional GRPO metrics to log. "
"`basic`: per-token ratio, KL, and advantage statistics. "
"`with_entropy`: also log per-token entropy. "
"Not supported with pipeline_parallel > 1."
),
hint=FieldHint.feature,
)

@property
def loss_class(self) -> "type[LanguageModelGRPOLoss]":
Expand Down
Loading
Loading