Thomas Viehmann
Thomas Viehmann
> you need to generate tokens up to the maximum length Well, so the rule is basically that the launch configuration of and parameters to the GPU kernel calls can't...
Some more detail: - make new symbols go to the right place, - how the user sets it. Given that we use the scopes, for the first, I'd probably push...
> This problem can also be properly resolved in prologue trace. i.e. here i1 is unpacked in prologue, because it is consumed by the top level symbol ltorch.getitem. Unfortunately the...
Related: - #1134 Issues for the steps: - #1222 - #1220 After 1220 is solved, we could use the present issue to track the remainder of the work. Inside the...
@KaelanDt
That seems rather unclear to me. I think we should not, possibly we should remove some subsymbols.
As an update, replacing `thunder_model` with ``` recipe = thunder.recipes.HFTransformers() recipe.executor_names = [ 'nvfuser', 'inplace_index_copy_ex', 'sdpa_mask_transform_ex', ] thunder_model = thunder.compile( model, recipe=recipe, # plugins=thunder.plugins.ReduceOverhead(), # CUDAGraphs will produce garbage output...
this is from torch.ops.higher_order.autograd_function_apply wanting #1134 (As mentioned elsewhere, I think a more timely way to fix would be to follow the torch.autograd.Function lookaside pattern and acquire the fw and...
Sure thing.
So this could be automated: https://github.com/Lightning-AI/lightning-thunder/blob/10a8a449109a2f33caa722b9cff2d588c8eb1954/thunder/core/jit_ext.py#L644-L652