minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[Auto Sync] Update test_deterministic.py (20260131) (#18034)

Browse Source 
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: Jiayi Yuan <34369239+jy-yuan@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
This commit is contained in:
Lianmin Zheng
2026-02-01 07:40:07 -08:00
committed by GitHub
co-authored by github-actions[bot] Jiayi Yuan gemini-code-assist[bot]
parent fb609669ca
commit 1acae30806
1 changed files with 20 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/test/test_deterministic.py
+20 -6
View File
@@ -320,6 +320,9 @@ class TokenIdsAndLogprobs:
token_ids: List[int] token_ids: List[int]
logprobs: List[float] logprobs: List[float]
# Logprob differences smaller than this are treated as non-divergent.
DIVERGENCE_EPS = 0.0
def __add__(self, other): def __add__(self, other):
return TokenIdsAndLogprobs( return TokenIdsAndLogprobs(
token_ids=self.token_ids + other.token_ids, token_ids=self.token_ids + other.token_ids,
@@ -381,13 +384,24 @@ class TokenIdsAndLogprobs:
# K3 approximation: KL(A||B) ≈ (exp(logr) - 1) - logr, where logr = log_a - log_b # K3 approximation: KL(A||B) ≈ (exp(logr) - 1) - logr, where logr = log_a - log_b
logr = logprobs_a - logprobs_b logr = logprobs_a - logprobs_b
kl_per_token = (np.exp(logr) - 1) - logr diverge_mask = np.abs(logr) > cls.DIVERGENCE_EPS
kl_mean = np.mean(kl_per_token) diverge_count = int(np.count_nonzero(diverge_mask))
kl_max = np.max(kl_per_token) total_count = int(logr.shape[0])
print(f" KL(A||B) mean: {kl_mean:.10e}") if diverge_count > 0:
print(f" KL(A||B) max : {kl_max:.10e}") kl_per_token = (np.exp(logr) - 1) - logr
print(f" Mean absolute logprob diff: {np.mean(np.abs(logr)):.10e}") kl_divergent = kl_per_token[diverge_mask]
kl_mean = float(np.mean(kl_divergent))
kl_max = float(np.max(kl_divergent))
mean_abs_logr = float(np.mean(np.abs(logr[diverge_mask])))
print(f" Divergent tokens: {diverge_count}/{total_count}")
print(f" KL(A||B) mean (divergent): {kl_mean:.10e}")
print(f" KL(A||B) max (divergent): {kl_max:.10e}")
print(
f" Mean absolute logprob diff (divergent): {mean_abs_logr:.10e}"
)
else:
print(f" Divergent tokens: 0/{total_count}")
return token_match and logprobs_match return token_match and logprobs_match