Infer llama2 vocab_size from tokenizer model when params.json provides vocab_size=-1

[l3utterfly]

🐛 Describe the bug

Following the instructions here: https://github.com/pytorch/executorch/tree/main/examples/models/llama2

I ran this command after downloading Llama2 weights: python3 -m examples.models.llama2.export_llama --checkpoint /path/to/Llama-2-7b/consolidated.00.pth --params /path/to/Llama-2-7b/params.json

I get this error: RuntimeError: Trying to create tensor with negative dimension -1: [-1, 4096]

Stacktrace:

INFO:datasets:PyTorch version 2.4.0.dev20240324+cpu available.
Could not import fairseq2 modules.
INFO:root:Loading model with checkpoint=/home/layla/src/text-generation-webui/models/Llama-2-7b/consolidated.00.pth, params=/home/layla/src/text-generation-webui/models/Llama-2-7b/params.json, use_kv_cache=False, weight_type=WeightType.LLAMA
Traceback (most recent call last):
  File "/home/layla/miniconda3/envs/executorch/lib/python3.10/runpy.py", line 196, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/home/layla/miniconda3/envs/executorch/lib/python3.10/runpy.py", line 86, in _run_code
    exec(code, run_globals)
  File "/home/layla/src/executorch/examples/models/llama2/export_llama.py", line 30, in <module>
    main()  # pragma: no cover
  File "/home/layla/src/executorch/examples/models/llama2/export_llama.py", line 26, in main
    export_llama(modelname, args)
  File "/home/layla/src/executorch/examples/models/llama2/export_llama_lib.py", line 504, in export_llama
    return _export_llama(modelname, args)
  File "/home/layla/src/executorch/examples/models/llama2/export_llama_lib.py", line 625, in _export_llama
    builder_exported_to_edge = _prepare_for_llama_export(
  File "/home/layla/src/executorch/examples/models/llama2/export_llama_lib.py", line 582, in _prepare_for_llama_export
    load_llama_model(
  File "/home/layla/src/executorch/examples/models/llama2/builder.py", line 83, in load_llama_model
    model, example_inputs, _ = EagerModelFactory.create_model(
  File "/home/layla/src/executorch/examples/models/model_factory.py", line 44, in create_model
    model = model_class(**kwargs)
  File "/home/layla/src/executorch/examples/models/llama2/model.py", line 139, in __init__
    self.model_ = Transformer(model_args)
  File "/home/layla/miniconda3/envs/executorch/lib/python3.10/site-packages/executorch/examples/models/llama2/llama_transformer.py", line 418, in __init__
    self.tok_embeddings = nn.Embedding(params.vocab_size, params.dim)
  File "/home/layla/miniconda3/envs/executorch/lib/python3.10/site-packages/torch/nn/modules/sparse.py", line 143, in __init__
    self.weight = Parameter(torch.empty((num_embeddings, embedding_dim), **factory_kwargs),
  File "/home/layla/miniconda3/envs/executorch/lib/python3.10/site-packages/torch/utils/_device.py", line 78, in __torch_function__
    return func(*args, **kwargs)
RuntimeError: Trying to create tensor with negative dimension -1: [-1, 4096]

Versions

CPU op-mode(s): 32-bit, 64-bit
Address sizes: 48 bits physical, 48 bits virtual
Byte Order: Little Endian
CPU(s): 32
On-line CPU(s) list: 0-31
Vendor ID: AuthenticAMD
Model name: AMD Ryzen Threadripper PRO 5955WX 16-Cores
CPU family: 25
Model: 8
Thread(s) per core: 2
Core(s) per socket: 16
Socket(s): 1
Stepping: 2
Frequency boost: enabled
CPU max MHz: 7031.2500
CPU min MHz: 1800.0000
BogoMIPS: 8000.05
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 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms 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 clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm
Virtualization: AMD-V
L1d cache: 512 KiB (16 instances)
L1i cache: 512 KiB (16 instances)
L2 cache: 8 MiB (16 instances)
L3 cache: 64 MiB (2 instances)
NUMA node(s): 1
NUMA node0 CPU(s): 0-31
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 Retbleed: Not affected
Vulnerability Spec rstack overflow: Mitigation; safe RET, no microcode
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp
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
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected

Versions of relevant libraries:
[pip3] executorch==0.1.0
[pip3] numpy==1.26.4
[pip3] torch==2.4.0.dev20240324+cpu
[pip3] torchao-nightly==2024.3.29
[pip3] torchaudio==2.2.0.dev20240324+cpu
[pip3] torchsr==1.0.4
[pip3] torchvision==0.19.0.dev20240324+cpu
[conda] executorch 0.1.0 pypi_0 pypi
[conda] numpy 1.26.4 pypi_0 pypi
[conda] torch 2.4.0.dev20240324+cpu pypi_0 pypi
[conda] torchao-nightly 2024.3.29 pypi_0 pypi
[conda] torchaudio 2.2.0.dev20240324+cpu pypi_0 pypi
[conda] torchsr 1.0.4 pypi_0 pypi
[conda] torchvision 0.19.0.dev20240324+cpu pypi_0 pypi

[dbort]

Thank you for reporting this issue @l3utterfly, and for all of the environment details!

Things are changing in this area pretty rapidly. Which specific git commit were you using when you saw this problem?

cc: @JacobSzwejbka @mikekgfb

[l3utterfly]

This is the commit hash I have in my environment: 57e3449

[dbort]

Thanks for the hash. What are the contents of your params.json file? I asked around, and one theory is that the vocab_size entry might be missing or might be -1. For llama7B, vocab_size should be 32000.

[l3utterfly]

Yes.. vocab size is "-1".

But this is from the official Llama2 repository on huggingface: https://huggingface.co/meta-llama/Llama-2-7b/blob/main/params.json

{"dim": 4096, "multiple_of": 256, "n_heads": 32, "n_layers": 32, "norm_eps": 1e-05, "vocab_size": -1}

Maybe we should update the documentation to add a line about needing to edit this manually?

[mikekgfb]

Or we check whether it is -1 and replace it with 32000, or the size of the accompanying tokenizer model? Either way we should generate a warning if/when we modify the params

[dbort]

@mikekgfb sounds like there are two steps here:

• Near term: Update the llama2 export docs to mention this problem and the workaround
• Longer term: If vocab_size is -1, infer the real size from the accompanying tokenizer model

[mergennachin]

#2926