Upload README.md with huggingface_hub

README.md

@@ -291,6 +291,7 @@ tokenizer = AutoTokenizer.from_pretrained(model_dir)
 model = AutoModelForCausalLM.from_pretrained(
     model_dir,
     dtype=torch.float32,
+    experts_implementation="batched_mm",
 ).to(device)
 model.eval()
 
@@ -339,6 +340,21 @@ The optional `flash-linear-attention` and `causal-conv1d` packages are not
 required. Without them, Transformers uses its PyTorch Gated DeltaNet fallback,
 which is the path used to train and validate this checkpoint.
 
+The examples explicitly select `experts_implementation="batched_mm"`. The
+model's routed experts have an intermediate width of 54, while the default
+PyTorch `grouped_mm` CUDA path in some recent Torch/Transformers combinations
+requires expert matrix strides to be multiples of 16 bytes. Without the
+explicit compatible implementation, loading succeeds but the first forward
+pass can fail with:
+
+```text
+RuntimeError: strides should be multiple of 16 bytes
+```
+
+`batched_mm` evaluates the same expert weights without that grouped-kernel
+layout restriction. `experts_implementation="eager"` is also a compatible,
+slower fallback.
+
 ## Intended uses
 
 This model is intended for: