ggml-cpu: add RISC-V Vector support for RWKV WKV6 operation

[ixgbe]

This PR adds RISC-V Vector (RVV) extension support for the RWKV WKV6 operation, enabling vectorized computation on RISC-V platforms.

[CISC]

Please note that ubuntu-cpu-cmake(-rpc)-riscv64-native will fail test-tokenizers-ggml-vocabs until git-lfs is fixed on the runners, all other tests will run to completion though.

[ggerganov]

I'm thinking that we should deprecate these ops since they are very model-specific:

llama.cpp/ggml/include/ggml.h

Lines 2389 to 2417 in 37adc9c

	GGML_API struct ggml_tensor * ggml_rwkv_wkv6(
	struct ggml_context * ctx,
	struct ggml_tensor * k,
	struct ggml_tensor * v,
	struct ggml_tensor * r,
	struct ggml_tensor * tf,
	struct ggml_tensor * td,
	struct ggml_tensor * state);
	GGML_API struct ggml_tensor * ggml_gated_linear_attn(
	struct ggml_context * ctx,
	struct ggml_tensor * k,
	struct ggml_tensor * v,
	struct ggml_tensor * q,
	struct ggml_tensor * g,
	struct ggml_tensor * state,
	float scale);
	GGML_API struct ggml_tensor * ggml_rwkv_wkv7(
	struct ggml_context * ctx,
	struct ggml_tensor * r,
	struct ggml_tensor * w,
	struct ggml_tensor * k,
	struct ggml_tensor * v,
	struct ggml_tensor * a,
	struct ggml_tensor * b,
	struct ggml_tensor * state);

Probably not worth investing much effort in optimizing. Rather, look to implement them as combination of other fundamental ops.

[ixgbe]

I'm thinking that we should deprecate these ops since they are very model-specific:

llama.cpp/ggml/include/ggml.h

Lines 2389 to 2417 in 37adc9c

	GGML_API struct ggml_tensor * ggml_rwkv_wkv6(
	struct ggml_context * ctx,
	struct ggml_tensor * k,
	struct ggml_tensor * v,
	struct ggml_tensor * r,
	struct ggml_tensor * tf,
	struct ggml_tensor * td,
	struct ggml_tensor * state);
	GGML_API struct ggml_tensor * ggml_gated_linear_attn(
	struct ggml_context * ctx,
	struct ggml_tensor * k,
	struct ggml_tensor * v,
	struct ggml_tensor * q,
	struct ggml_tensor * g,
	struct ggml_tensor * state,
	float scale);
	GGML_API struct ggml_tensor * ggml_rwkv_wkv7(
	struct ggml_context * ctx,
	struct ggml_tensor * r,
	struct ggml_tensor * w,
	struct ggml_tensor * k,
	struct ggml_tensor * v,
	struct ggml_tensor * a,
	struct ggml_tensor * b,
	struct ggml_tensor * state);

Probably not worth investing much effort in optimizing. Rather, look to implement them as combination of other fundamental ops.

Wanted to check in on this PR. Should I:

• Wait for further discussion on the WKV op deprecation?
• Make any changes to the current implementation?

Happy to follow whatever direction works best for the project. Thanks!

[ggerganov]

The best way is to try to implement these operations with other ggml ops. If we can do that, then we can replace the current implementations with the composite ones and make the deprecation process smoother. The main question is if they can be expressed as composite ops in a meaningful way.