Fix MHAEinsum weight dimension bug when d_in != d_out (#857) (#893)

* Fix MHAEinsum weight dimension bug when d_in != d_out (#857)

Previously MHAEinsum initialized weight matrices with shape (d_out, d_in) and used inappropriate einsum notation, causing failures for non-square input-output dimensions. This commit corrects weight initialization to shape (d_in, d_out), updates einsum notation to 'bnd,do->bno', and adds three unit tests to verify parity across different d_in and d_out settings. All tests pass successfully.

* use pytest

* Update .gitignore

---------

Co-authored-by: rasbt <[email protected]>

Author: aviralgarg05

.github/workflows/basic-tests-linux-uv.yml

@@ -48,6 +48,7 @@ jobs:
         run: |
           source .venv/bin/activate
           pytest setup/02_installing-python-libraries/tests.py
+          pytest ch03/02_bonus_efficient-multihead-attention/tests/test_mha_implementations.py
           pytest ch04/01_main-chapter-code/tests.py
           pytest ch04/03_kv-cache/tests.py
           pytest ch05/01_main-chapter-code/tests.py

.gitignore

@@ -328,3 +328,5 @@ cython_debug/
 # pixi environments
 .pixi
 *.egg-info
+
+

ch03/02_bonus_efficient-multihead-attention/mha-implementations.ipynb

@@ -0,0 +1,63 @@
+from pathlib import Path
+import torch
+import pytest
+
+
+from llms_from_scratch.utils import import_definitions_from_notebook
+
+
+@pytest.fixture
+def nb_imports():
+    nb_dir = Path(__file__).resolve().parents[1]
+    mod = import_definitions_from_notebook(nb_dir, "mha-implementations.ipynb")
+    return mod
+
+
+def copy_weights(from_mha, to_mha):
+    with torch.no_grad():
+        to_mha.W_query.copy_(from_mha.W_query.weight.T)
+        to_mha.W_key.copy_(from_mha.W_key.weight.T)
+        to_mha.W_value.copy_(from_mha.W_value.weight.T)
+
+        to_mha.out_proj.weight.copy_(from_mha.out_proj.weight)
+        to_mha.out_proj.bias.copy_(from_mha.out_proj.bias)
+
+
+@pytest.mark.parametrize(
+    "d_in,d_out,batch,seq_len,num_heads,seed",
+    [
+        (768, 768, 2, 4, 12, 123),  # d_in == d_out
+        (768, 1536, 2, 4, 12, 456),  # d_in != d_out
+        (1024, 512, 2, 4, 8, 789),   # d_in > d_out
+    ],
+)
+def test_mha_einsum_matches_ch03(d_in, d_out, batch, seq_len, num_heads, seed, nb_imports):
+    torch.manual_seed(seed)
+
+    x = torch.randn(batch, seq_len, d_in)
+
+    mha_linear = nb_imports.Ch03_MHA(
+        d_in=d_in,
+        d_out=d_out,
+        context_length=seq_len,
+        dropout=0.0,
+        num_heads=num_heads,
+        qkv_bias=False,
+    ).eval()
+
+    mha_einsum = nb_imports.MHAEinsum(
+        d_in=d_in,
+        d_out=d_out,
+        context_length=seq_len,
+        dropout=0.0,
+        num_heads=num_heads,
+        qkv_bias=False,
+    ).eval()
+
+    copy_weights(mha_linear, mha_einsum)
+
+    out_linear = mha_linear(x)
+    out_einsum = mha_einsum(x)
+
+    assert out_linear.shape == out_einsum.shape == torch.Size([batch, seq_len, d_out])
+    assert torch.allclose(out_linear, out_einsum, atol=1e-5)