Llama 3.2 1B export appears to be broken on Executorch Vulkan backend #16647
Description
🐛 Describe the bug
I am trying to export Llama 3.2 1B to run on an android device using the Vulkan backend, and was following this documentation: https://docs.pytorch.org/executorch/main/backends/vulkan/tutorials/etvk-llama-tutorial.html. I was unable to install the executorch package from pip for some reason, so I built it from source following this: https://docs.pytorch.org/executorch/main/using-executorch-building-from-source.html. After that, for exporting Llama I run the command:
python -m examples.models.llama.export_llama \
-c $HOME/.llama/checkpoints/${LLM_NAME}-${LLM_SIZE}${LLM_SUFFIX}/consolidated.00.pth \
-p $HOME/.llama/checkpoints/${LLM_NAME}-${LLM_SIZE}${LLM_SUFFIX}/params.json \
-d fp32 --${BACKEND} \
-qmode ${QUANT} -G ${GROUP_SIZE} \
--max_seq_length ${CONTEXT_LENGTH} \
--max_context_length ${CONTEXT_LENGTH} \
-kv --use_sdpa_with_kv_cache \
--metadata '{"append_eos_to_prompt": 0, "get_bos_id":128000, "get_eos_ids":[128009, 128001]}' \
--model "llama3_2" \
--output_name $HOME/.llama/checkpoints/${LLM_NAME}-${LLM_SIZE}${LLM_SUFFIX}/${LLM_NAME}-${LLM_SIZE}${LLM_SUFFIX}_${BACKEND}_${QUANT}_g${GROUP_SIZE}_c${CONTEXT_LENGTH}.pte
and it fails with a dynamic shape error in the slice_copy operation. The main segment of the error is this:
While executing %aten_slice_copy_tensor : [num_users=2] = call_function[target=executorch.exir.dialects.edge._ops.aten.slice_copy.Tensor](args = (%b_rope_freqs_cos, 0, %_local_scalar_dense, %add), kwargs = {})
Original traceback:
File "/users/Lakshman/executorch/.executorch-qnn/lib/python3.12/site-packages/torch/_dynamo/functional_export.py", line 216, in forward
res = self._export_root(*args, **kwargs)
File "/users/Lakshman/executorch/.executorch-qnn/lib/python3.12/site-packages/executorch/examples/models/llama/llama_transformer.py", line 196, in forward
freqs_cos, freqs_sin = self.rope.get_freqs(
File "/users/Lakshman/executorch/.executorch-qnn/lib/python3.12/site-packages/executorch/examples/models/llama/rope.py", line 300, in get_freqs
freqs_cos = self.freqs_cos.narrow(0, input_pos_item, seq_len)
Use tlparse to see full graph. (https://github.com/pytorch/tlparse?tab=readme-ov-file#tlparse-parse-structured-pt2-logs)
Since the error seems to be with dynamic shapes I ran the command again with the --disable_dynamic_shapes flag to bypass the failing section of the code, but then the build fails with a KeyError in get_op_features in the op_registry.py file:
File "/users/Lakshman/executorch/.executorch-qnn/lib/python3.12/site-packages/executorch/backends/vulkan/op_registry.py", line 877, in get_op_features
return vulkan_supported_ops[target.name()]
~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^
KeyError: 'aten::index.Tensor'
Which I think means the index tensor operation isn't a supported Vulkan Op but is used somewhere? I tried disabling quantization and running the build, but I run into the same error. I patched that by returning a basic default OpFeatures object if the operation is not in vulkan_supported_ops. Following this I encountered an SDPA error, which I think expects KV cache update nodes, which the Vulkan partitioner seems to explicitly skip (all the llama.update_cache.default nodes were skipped). Is this expected?
I finally tried the export without the SPDA or KV cache flags, and then it seems to finally pass and export.
Versions
I want to know if I am missing some other dependencies/steps that may not be explicitly mentioned in the tutorial that fixes all these issues much more directly, since the current version I have working definitely does not seem to be the intended way to export the model. For building from source, the default android ndk glslc also does not work as noted here: #14507 and so I run the build with the manually installed version of the vulkan SDK which I don't think is mentioned as a requirement in the building from source docs.
I am using the Vulkan SDK version 1.4.335 and this is the output of collect env. I am using venv, not conda.
Collecting environment information...
PyTorch version: 2.11.0.dev20251222+cpu
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A
OS: Ubuntu 24.04.3 LTS (x86_64)
GCC version: (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
Clang version: Could not collect
CMake version: version 3.31.10
Libc version: glibc-2.39
Python version: 3.12.3 (main, Jan 8 2026, 11:30:50) [GCC 13.3.0] (64-bit runtime)
Python platform: Linux-6.8.0-71-generic-x86_64-with-glibc2.39
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
Is XPU available: False
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
Caching allocator config: N/A
CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 43 bits physical, 48 bits virtual
Byte Order: Little Endian
CPU(s): 128
On-line CPU(s) list: 0-127
Vendor ID: AuthenticAMD
Model name: AMD EPYC 7543 32-Core Processor
CPU family: 25
Model: 1
Thread(s) per core: 2
Core(s) per socket: 32
Socket(s): 2
Stepping: 1
BogoMIPS: 5590.00
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local user_shstk clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin brs arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca sev sev_es debug_swap
Virtualization: AMD-V
L1d cache: 2 MiB (64 instances)
L1i cache: 2 MiB (64 instances)
L2 cache: 32 MiB (64 instances)
L3 cache: 512 MiB (16 instances)
NUMA node(s): 8
NUMA node0 CPU(s): 0-7,64-71
NUMA node1 CPU(s): 8-15,72-79
NUMA node2 CPU(s): 16-23,80-87
NUMA node3 CPU(s): 24-31,88-95
NUMA node4 CPU(s): 32-39,96-103
NUMA node5 CPU(s): 40-47,104-111
NUMA node6 CPU(s): 48-55,112-119
NUMA node7 CPU(s): 56-63,120-127
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed: Not affected
Vulnerability Spec rstack overflow: Mitigation; Safe RET
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Retpolines; IBPB conditional; IBRS_FW; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
Versions of relevant libraries:
[pip3] executorch==1.1.0a0+082b62b
[pip3] numpy==2.4.1
[pip3] pytorch_tokenizers==1.0.1
[pip3] torch==2.11.0.dev20251222+cpu
[pip3] torchao==0.16.0+git08e5e203f
[pip3] torchaudio==2.10.0.dev20251222+cpu
[pip3] torchdata==0.11.0
[pip3] torchsr==1.0.4
[pip3] torchtune==0.6.1
[pip3] torchvision==0.25.0.dev20251222+cpu
[conda] Could not collect