diff --git a/docs/ai/encoder_cpu_wte_backward_bucket_width.md b/docs/ai/encoder_cpu_wte_backward_bucket_width.md
new file mode 100644
index 0000000..08b2943
--- /dev/null
+++ b/docs/ai/encoder_cpu_wte_backward_bucket_width.md
@@ -0,0 +1,72 @@
+# Encoder CPU `dwte` backward skipped token rows
+
+**Symptom.** `make test-python` failed on all three
+`tests/test_encoder_equivalence.py::test_backward_matches_torch` cases. The
+position-embedding gradient (`dwpe`) matched PyTorch, but the token-embedding
+gradient (`dwte`) stayed entirely zero for rows whose tokens appeared in the
+batch.
+
+For the small fp32 case, PyTorch expected non-zero gradients for token rows
+`[5, 12, 15, 16, 25, 26, 27, 28]`; the custom op returned no non-zero `dwte`
+rows at all.
+
+## Cause
+
+The CPU `wte_backward_cpu` path derived each bucket's channel span with:
+
+```mojo
+var c_per_warp = WARP_SIZE * width
+```
+
+That is the right shape for GPU terminology, but the CPU path is consuming
+bucket metadata produced with the module-level bucket contract:
+
+```mojo
+comptime WTE_C_PER_WARP = 32 * WTE_BWD_SIMD_WIDTH
+```
+
+On this MAX/Mojo CPU custom-op path, the mismatch made `c_len` evaluate to zero
+for the Python equivalence cases, so the kernel never entered the accumulation
+loop that writes `dwte`.
+
+## Fix
+
+Use `WTE_C_PER_WARP` directly in the CPU backward path when mapping a bucket's
+`channel_group` to `[c_base, c_end)`. This keeps the consumer in lockstep with
+`build_wte_buckets` and the Python test helper, both of which emit bucket groups
+using the same constant.
+
+## Verification
+
+Commands run on Apple Silicon/macOS with Pixi-installed Mojo
+`1.0.0b3.dev2026072606`:
+
+```bash
+PATH=/Users/saisandeepkantareddy/.pixi/bin:$PATH \
+LLMM_DISABLE_MEF_CACHE=1 \
+/Users/saisandeepkantareddy/.pixi/bin/pixi -q run pytest -q \
+  tests/test_encoder_equivalence.py
+# 6 passed
+
+PATH=/Users/saisandeepkantareddy/.pixi/bin:$PATH \
+/Users/saisandeepkantareddy/.pixi/bin/pixi -q run pytest -q \
+  tests/test_dataloader.py \
+  tests/test_tokenizer_equivalence.py \
+  tests/test_encoder_equivalence.py
+# 12 passed
+
+PATH=/Users/saisandeepkantareddy/.pixi/bin:$PATH \
+make PIXI='/Users/saisandeepkantareddy/.pixi/bin/pixi -q' test-python
+# 233 passed, 15 skipped
+
+PATH=/Users/saisandeepkantareddy/.pixi/bin:$PATH \
+make PIXI='/Users/saisandeepkantareddy/.pixi/bin/pixi -q' build
+# passed
+```
+
+The Mojo build printed Crashpad warnings under the sandboxed macOS environment,
+but exited successfully and produced `build/train_gpt2`.
+
+## AI use statement
+
+Written with AI assistance (Codex/Agent), directed by Sai Sandeep Kantareddy.
diff --git a/llmm/encoder.mojo b/llmm/encoder.mojo
index 7a6f81c..205d881 100644
--- a/llmm/encoder.mojo
+++ b/llmm/encoder.mojo
@@ -651,10 +651,9 @@ def wte_backward_cpu[
             var token_idx = Int(info[2])
             var channel_group = Int(info[3])
 
-            var c_per_warp = WARP_SIZE * width
-            var c_base = channel_group * c_per_warp
+            var c_base = channel_group * WTE_C_PER_WARP
 
-            var c_end = min(c_base + c_per_warp, channels)
+            var c_end = min(c_base + WTE_C_PER_WARP, channels)
             var c_len = c_end - c_base
 
             if c_len > 0: