Get Falcon 7B running E2E

[lanchongyizu]

Ticket

#1045

Problem description

See the details in the above ticket.

What's changed

* Add disable_load_params_once option for model testing since Falcon 7B
  is incompatible with load_params_once which causes OOM.
* Convert following torch ops to ttnn ops:
    {
      "op_name": "aten.squeeze.default",
      "op_schema": "Tensor<[1, 7]> self = ?"
    },
    {
      "op_name": "aten.native_layer_norm.default",
      "op_schema": "Tensor<[1, 7, 4544]> input = ?,<br>List[int] normalized_shape = [4544],<br>Optional[Tensor]<[4544]> weight = ?,<br>Optional[Tensor]<[4544]> bias = ?,<br>float eps = 1e-05"
    },
    {
      "op_name": "aten.index.Tensor",
      "op_schema": "Tensor<[1, 7, 73, 64]> self = ?,<br>List[Optional[Tensor]] indices = [None, None, _folded_lift_fresh_copy]"
    },
    {
      "op_name": "aten.mul.Scalar",
      "op_schema": "Tensor<[1, 71, 7, 64]> self = ?,<br>number other = 0.3535533905932738"
    },
* Enable batch_size 32 for Falcon 7B E2E testing

OOM issue:
As mentioned by @kevinwuTT in another issue, Falcon 7B is incompatible with load_params_once. After load_params_once is disabled, OOM issue disappears for both batchsize 1 and 32.

Model	Status	Batch	Compiled First Run (ms)	Original Throughput (Inferences Per Second)	Compiled Throughput (Inferences Per Second)	Accuracy (%)	Torch Ops Before (Unique Ops)	Torch Ops Remain (Unique Ops)	To/From Device Ops
Falcon-7B	✅	1	44009.8	0.384783	0.13878	98	2600 (27)	0 (0)	193
Falcon-7B	✅	32	50098.8	5.07062	4.27947	98	2696 (27)	0 (0)	193

[marty1885]

Ping. @ayerofieiev-tt if you can get someone to look at this one

[jmalone-tt]

These changes look reasonable. Happy to merge if tests pass (kicked off a run here: https://github.com/tenstorrent/pytorch2.0_ttnn/actions/runs/19867910005)

[jmalone-tt]

Looks like it still hits OOM in the latest run. Have you verified that it runs on your end?
https://github.com/tenstorrent/pytorch2.0_ttnn/actions/runs/19867910005/job/57557971201

[lanchongyizu]

Looks like it still hits OOM in the latest run. Have you verified that it runs on your end? https://github.com/tenstorrent/pytorch2.0_ttnn/actions/runs/19867910005/job/57557971201

Sure. As mentioned in #1265 (comment), Falcon 7B is incompatible with load_params_once, so please add --disable_load_params_once in the pytest command line as below.

python3 -m pytest --github-report tests/models/falcon/test_falcon.py --disable_load_params_once --report_nth_iteration=$num_iterations --export_code=accuracy --splits 57 --group 10 -s

[jmalone-tt]

We just merged another PR that was hitting the same issue - if you look at tests/models/mamba/test_mamba, you should see a new pytest fixture load_params_once. Please rebase your changes and add that fixture, and if tests pass we should be good to merge