happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

removed all traces of kda_scan

This commit is contained in:
Yee Man Chan 2025-12-19 07:36:06 +08:00
parent f9a11d7758
commit 776294c04e
4 changed files with 0 additions and 30 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
ggml/include/ggml.h
View File

@ -539,7 +539,6 @@ extern "C" {
GGML_OP_FLASH_ATTN_BACK,
GGML_OP_SSM_CONV,
GGML_OP_SSM_SCAN,
GGML_OP_KDA_SCAN,
GGML_OP_WIN_PART,
GGML_OP_WIN_UNPART,
GGML_OP_GET_REL_POS,
@ -2338,28 +2337,6 @@ extern "C" {
struct ggml_tensor * C,
struct ggml_tensor * ids);
// KDA (Kimi Delta Attention) scan
// Delta attention recurrence:
// h[t] = exp(g[t]) * h[t-1] + k[t]^T * (beta[t] * (v[t] - h[t-1] @ k[t]))
// o[t] = q[t]^T @ h[t]
// Parameters:
// h: hidden state {head_dim, head_dim, n_head, n_seqs+}
// q: query {head_dim, n_head, n_seq_tokens, n_seqs}
// k: key {head_dim, n_head, n_seq_tokens, n_seqs}
// v: value {head_dim, n_head, n_seq_tokens, n_seqs}
// g: gate {head_dim, n_head, n_seq_tokens, n_seqs}
// beta: mixing {n_head, n_seq_tokens, n_seqs}
// ids: seq indices {n_seqs}
GGML_API struct ggml_tensor * ggml_kda_scan(
struct ggml_context * ctx,
struct ggml_tensor * h,
struct ggml_tensor * q,
struct ggml_tensor * k,
struct ggml_tensor * v,
struct ggml_tensor * g,
struct ggml_tensor * beta,
struct ggml_tensor * ids);
// partition into non-overlapping windows with padding if needed
// example:
// a: 768 64 64 1

1
ggml/src/ggml-cpu/ggml-cpu.c
View File

@ -2320,7 +2320,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
case GGML_OP_FLASH_ATTN_BACK:
case GGML_OP_SSM_CONV:
case GGML_OP_SSM_SCAN:
case GGML_OP_KDA_SCAN:
case GGML_OP_RWKV_WKV6:
case GGML_OP_GATED_LINEAR_ATTN:
case GGML_OP_RWKV_WKV7:

5
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@ -4503,11 +4503,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
return op->src[0]->ne[0] == 16 && op->src[0]->ne[1] == 1 && op->src[0]->ne[2] % 128 == 0 && op->src[4]->ne[1] == 1;
}
}
case GGML_OP_KDA_SCAN: {
// KDA scan kernel supports head_dim 64 or 128
const int64_t head_dim = op->src[0]->ne[0];
return head_dim == 64 || head_dim == 128;
}
case GGML_OP_SSM_CONV: {
// assumes d_inner % threads == 0
return op->src[0]->ne[1] % 128 == 0;

1
ggml/src/ggml.c
View File

@ -999,7 +999,6 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"FLASH_ATTN_BACK",
"SSM_CONV",
"SSM_SCAN",
"KDA_SCAN",
"WIN_PART",
"WIN_UNPART",
"GET_REL_POS",