Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

GRPO kept-set refusals: stop saying UNSLOTH_GRPO_TRUNCATED_SUPPORT=1 was set (#15307)

Daniel Hanfbc905fd02a9
diff --git a/tests/test_grpo_truncated_support_policy_gate.py b/tests/test_grpo_truncated_support_policy_gate.py
index c0dba35cdc1..8c262d100fd 100644
--- a/tests/test_grpo_truncated_support_policy_gate.py
+++ b/tests/test_grpo_truncated_support_policy_gate.py
@@ -161,6 +161,26 @@ def test_policies_with_their_own_kept_set_score_the_full_vocabulary(monkeypatch,
assert not messages
+def test_refusals_do_not_claim_the_variable_was_set(monkeypatch):
+ # zoo's automatic gate means refusals fire with the variable unset: never say "ignoring ...=1".
+ monkeypatch.delenv("UNSLOTH_GRPO_TRUNCATED_SUPPORT", raising = False)
+ _, messages = _bind(monkeypatch, _trainer(beta = 0.04))
+ refuse = _refuse_helper(monkeypatch, _SUPPORT)
+ trainer = _trainer()
+ trainer._unsloth_vllm_sync_skip = True
+ with warnings.catch_warnings(record = True) as caught:
+ warnings.simplefilter("always")
+ refuse(trainer)
+ messages += [str(w.message) for w in caught]
+ assert len(messages) == 2
+ for message in messages:
+ assert "ignoring" not in message
+ assert message.startswith(
+ "Unsloth: not scoring over the rollouts' top_k / top_p / min_p kept set"
+ )
+ assert "automatic" in message
+
+
def _own_support_logps(logits, index, top_k):
"""The per-policy rule: each policy's own top-k set, plus the scored token."""
kth = torch.topk(logits, top_k, dim = -1).values[:, -1:]
diff --git a/unsloth/models/rl_replacements.py b/unsloth/models/rl_replacements.py
index 54f2032995b..c8f354f8834 100644
--- a/unsloth/models/rl_replacements.py
+++ b/unsloth/models/rl_replacements.py
@@ -1611,7 +1611,9 @@ def _unsloth_grpo_refuse_unscored_support(self):
if mismatch:
import warnings
warnings.warn(
- "Unsloth: ignoring UNSLOTH_GRPO_TRUNCATED_SUPPORT=1 because "
+ "Unsloth: not scoring over the rollouts' top_k / top_p / min_p kept set "
+ "(UNSLOTH_GRPO_TRUNCATED_SUPPORT=1, or automatic for truncated vLLM rollouts "
+ "feeding the IS correction or off-policy mask) because "
f"{', '.join(mismatch)}: the scored policy is not the sampler, so its top_k / top_p / "
"min_p kept set differs from vLLM's and the PPO ratio would be unnormalised. "
"Scoring over the full vocabulary instead.",
@@ -1621,7 +1623,7 @@ def _unsloth_grpo_refuse_unscored_support(self):
def _unsloth_grpo_support_mismatch(self):
- """Why UNSLOTH_GRPO_TRUNCATED_SUPPORT=1 cannot score this trainer, or [] when it can.
+ """Why kept-set scoring (UNSLOTH_GRPO_TRUNCATED_SUPPORT) cannot score this trainer, or [] when it can.
Each scoring call rebuilds the top_k / top_p / min_p kept set from its own logits and always keeps the
scored token. That is the sampler's set only while the scored weights are the rollout's. A reference
@@ -3543,8 +3545,8 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
**kwargs,
)
- # Opt-in (UNSLOTH_GRPO_TRUNCATED_SUPPORT=1): score over the rollout's top_k / top_p /
- # min_p kept set, matching vLLM's processed_logprobs. The zoo loss refuses to mix
+ # UNSLOTH_GRPO_TRUNCATED_SUPPORT (=1 or zoo's automatic gate): score over the rollout's
+ # top_k / top_p / min_p kept set, matching vLLM's processed_logprobs. The zoo loss refuses to mix
# normalisations unless this marker says the old / ref logps were scored the same way.
try:
from unsloth_zoo.rl_replacements import grpo_truncated_support_kwargs
@@ -3557,7 +3559,9 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
if not getattr(self, "_unsloth_truncated_support_refused", False):
import warnings
warnings.warn(
- "Unsloth: ignoring UNSLOTH_GRPO_TRUNCATED_SUPPORT=1 because "
+ "Unsloth: not scoring over the rollouts' top_k / top_p / min_p kept set "
+ "(UNSLOTH_GRPO_TRUNCATED_SUPPORT=1, or automatic for truncated vLLM "
+ "rollouts feeding the IS correction or off-policy mask) because "
f"{', '.join(_support_mismatch)}: those logprobs come from weights other "
"than the rollout's, whose top_k / top_p / min_p kept set differs from the "
"sampler's, so per-policy kept sets give unnormalised PPO ratios and KL. "