Eugene Kliuchnikov

Results 156 comments of Eugene Kliuchnikov

Should be fixed by #4007

I've pinpointed the problem. Effort affects the value `splitting_heuristics_node_threshold`. Higher efforts make it smaller... But around the sweet spot smaller does always not mean better. (And anyways, effort-11 does not...

(does not build even without LTO...)

I'm open to adding more options; will prepare PR soon.

Failure is reported by `ValidateChannelDimensions`. Size threshold depends on "group dim" that depends on "group shift".

Disabling this check causes considerable diff in "streaming_fuzzer" (-> failure).

@jonsneyers could you take a look at `ValidateChannelDimensions` / `DefaultSqueezeParameters`. Likely one is over-restrictive or the other is not taking into account group dim...

Re-open; a case found where `ValidateChannelDimensions` fails.