[GPT-OSS-120B] Reference implementation #2395
Conversation
add better parsing and check for harmony tokens
|
|
||
|
|
||
| def decode_hex_to_tokens(hex_data: str) -> List[int]: | ||
| """Decode hex string to list of token IDs (int32). |
There was a problem hiding this comment.
not a concern here - we added int32 after int64 didnt work so ds-r1 code has support for both while defaulting to int32:
|
|
||
| total_correct = sum(stats["correct"] for stats in dataset_stats.values()) | ||
| total_samples = sum(stats["total"] for stats in dataset_stats.values()) | ||
| overall_accuracy = ( |
There was a problem hiding this comment.
consider putting formulas on one line, if block on next
similar below (final-score-percentage)
There was a problem hiding this comment.
IIUC these are linux soft links?
There was a problem hiding this comment.
yeah that is right
|
@arjunsuresh @mrmhodak can we try to get this merged soon? 🙏 |
|
@v-shobhit once the WG approves I can merge it. I'm told to not do any merge outside of the WG discussions 😕 |
hanyunfan
dismissed
nvzhihanj’s stale review
Zhihan confirmed this isn’t an issue and will address it after the merge, if needed.
|
@anandhu-eng will create an empty commit as a workaround for this checker issue. This problem ultimately needs to be resolved by the Automation Work Task Force. |
|
PR is ready to be merged once approved. |
hanyunfan
left a comment
hanyunfan
left a comment
There was a problem hiding this comment.
As we talked in WG, this is approved.
|
@v-shobhit merged. thanks for all the hard work! |
10 participants
Adding gpt-oss-120b reference implementation.
This uses SGLang backend to serve gpt-oss-120b model. Scripts are provided to run this in Offline/Server scenario in PerformanceOnly/AccuracyOnly modes.