Question about k_grouped_gemm #243
Description
Why, when computing wgrad in k_grouped, do dy and dx use per_channel_cast_to_fp8 instead of per_token_cast_to_fp8?
{{ message }}
Why, when computing wgrad in k_grouped, do dy and dx use per_channel_cast_to_fp8 instead of per_token_cast_to_fp8?