Lightning-AI lightning-thunder issues

Lightning-AI / lightning-thunder

Make PyTorch models up to 40% faster! Thunder is a source to source compiler for PyTorch. It enables using different hardware executors at once; across one or thousands of GPUs.

Apache License 2.0

1.15k stars 77 forks source link

issues

Newest

Newest Most commented Recently updated Oldest Least commented Least recently updated

Adding shape prim

#1113 jjsjann123 closed 2 weeks ago
12
Remove redundant parametrizations for swiglu benchmark

#1112 IvanYashchuk closed 3 weeks ago
0
quantization: process tensors on meta device directly, maybe implement CPU quantization (if it is easy)

#1111 t-vi opened 3 weeks ago
4
patching inplace update nvfuser translation

#1110 jjsjann123 closed 3 weeks ago
3
nvfuser translation with inplace update is not correct

#1109 jjsjann123 closed 3 weeks ago
0
[benchmark_litgpt] always `destroy_process_group` with `try-except-finally`

#1108 crcrpar closed 3 weeks ago
0
try except as some `untyped_storage` is invalid for `data_ptr`, e.g. floa…

#1107 crcrpar opened 3 weeks ago
0
Define `torchsymbol` for `torch.ops.higher_order.autograd_function_apply`

#1106 crcrpar closed 2 days ago
8
Add a parametrized benchmark for swiglu

#1105 IvanYashchuk closed 3 weeks ago
2
pickling failure for auto registered ops fix

#1104 k223kim closed 2 weeks ago
2
Add liger kernels as an executor

#1103 mruberry opened 3 weeks ago
0
Use the proxy args in the `FSDPParamUnpaddingVisitor`

#1102 kiya00 closed 3 weeks ago
3
support list[int] as shape for torch.randn

#1101 kshitij12345 closed 3 weeks ago
0
switch ci jobs to 2.4

#1100 t-vi closed 3 weeks ago
0
Transform to skip distributed collective ops when `group.size() == 1`

#1099 crcrpar opened 3 weeks ago
0
VJP rule for `PrimIDs.CLONE`, with some `torch.memory_format` handling logic

#1098 crcrpar closed 2 weeks ago
0
Register `PrimIDs.BITWISE_*` to `nondifferentiable_vjp_symbols`

#1097 crcrpar closed 3 weeks ago
0
nvfuserex to leverage static shape in trace

#1096 jjsjann123 closed 2 weeks ago
1
Add allow_cpu_scalar_tensors for shape,device check and broadcast functions

#1095 kiya00 closed 2 weeks ago
13
thunder + dynamo : use recompile instead of real_recompile

#1094 kshitij12345 closed 2 weeks ago
3
Expected dtype thunder.dtypes.bfloat16 but found thunder.dtypes.float32 for Dynamo+Thunder and Mixtral-8x7B-v0.1

#1093 mpatel31415 closed 2 weeks ago
8
unhashable type: slice for Thunder and Nous-Hermes-13b

#1092 mpatel31415 closed 3 weeks ago
0
AttributeError for Dynamo+Thunder+FP8 and Phi-3-mini-4k-instruct 'GraphModule' object has no attribute 'real_recompile'

#1091 mpatel31415 closed 2 weeks ago
1
Remove thunder+nvfuser+torch.compile executor from default list in targets.py

#1090 IvanYashchuk closed 3 weeks ago
0
[CI Fix] Update check-package to v0.11.7

#1089 shino16 closed 3 weeks ago
0
bump pytorch/triton to 2.4/3.0.0

#1088 crcrpar closed 3 weeks ago
1
bring back `IN_PLACE` tag to copy

#1087 crcrpar closed 3 weeks ago
0
After adding recomputing_symbols, sort the new consumer according to the data flow

#1086 kiya00 closed 3 weeks ago
1
`masked_fill` casts `value` to `a.dtype`

#1085 crcrpar closed 3 weeks ago
0
Allow different device and dtype for `prims.copy_` args

#1084 crcrpar opened 4 weeks ago
0
`thunder.jit`ted `Tensor.masked_fill` and `Tensor.masked_fill_` return `torcdh.float32` tensor even when input is `torch.int64`

#1083 crcrpar closed 3 weeks ago
0
Add option to `ThunderCompiler` to save `gm.code` or `gm.print_readable` to file

#1082 crcrpar opened 4 weeks ago
1
Bump pypa/gh-action-pypi-publish from 1.9.0 to 1.10.0

#1081 dependabot[bot] closed 3 weeks ago
0
Bump nbsphinx from 0.9.4 to 0.9.5

#1080 dependabot[bot] closed 3 weeks ago
0
Bump myst-parser from 1.0.0 to 4.0.0

#1079 dependabot[bot] closed 4 weeks ago
2
Bump litgpt from 0.3.1 to 0.4.11

#1078 dependabot[bot] closed 3 weeks ago
1
Bump ipython[all] from 8.25.0 to 8.27.0

#1077 dependabot[bot] closed 3 weeks ago
0
Bump hypothesis from 6.104.2 to 6.111.2

#1076 dependabot[bot] closed 3 weeks ago
0
thunder as torch.compile backend - Add `splitter` to pass unsupported regions to inductor

#1075 kshitij12345 closed 2 weeks ago
9
Memory Leak when using nn.Module hooks and thunder.jit

#1074 kshitij12345 closed 6 days ago
2
use `torch.OP##_` if available not `torch.Tensor.OP##_`

#1073 crcrpar closed 3 weeks ago
0
`torch.exp_` is not suppurted, but `torch.Tensor.exp_` is

#1072 crcrpar closed 3 weeks ago
0
trace: add cursor to for bsym insertion

#1071 t-vi opened 1 month ago
1
Add `TensorProxy.grad` attribute and proxify `Tensor.grad`

#1070 shino16 closed 3 days ago
1
revert using transform to execution, add ad-hoc fix for type_as

#1069 t-vi closed 1 month ago
1
Implement VJP for Clone

#1068 tfogal closed 2 weeks ago
2
Proxy order KeyError fix

#1067 riccardofelluga opened 1 month ago
7
Add graph-by-graph benchmarking of dynamo.ThunderCompiler

#1066 kiya00 opened 1 month ago
1
torch_compile_cat_ex doesn't work when an input is a registered buffer on the module

#1065 IvanYashchuk closed 1 month ago
5
Add a bunch of benchmarks from NeMo's NeVA

#1064 tfogal closed 6 days ago
10

Previous Next