highperfocused/ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2025-11-12 09:57:51 +01:00
Code Issues Packages Projects Releases Wiki Activity

Update GGML to b6646 (#12245)

Browse Source 
Notable EOLs with this change:
- MacOS v12 and v13 are no longer supported (v14+ required)
- AMD gfx900 and gfx906 are no longer supported
This commit is contained in:
Daniel Hiltgen
2025-10-02 14:47:10 -07:00
committed by GitHub
parent fdb109469f
commit c68f367ef6
326 changed files with 30615 additions and 20624 deletions
Download Patch File Download Diff File Expand all files Collapse all files

94
llama/llama.cpp/src/llama-graph.h vendored
View File

@@ -19,8 +19,8 @@ struct llama_cparams;
struct llama_memory_context_i;
class llama_kv_cache_unified_context;
class llama_kv_cache_unified_iswa_context;
class llama_kv_cache_context;
class llama_kv_cache_iswa_context;
class llama_memory_recurrent_context;
class llama_memory_hybrid_context;
@@ -78,6 +78,11 @@ struct llm_graph_params;
class llm_graph_input_i {
public:
llm_graph_input_i() {
const char * LLAMA_GRAPH_INPUT_DEBUG = getenv("LLAMA_GRAPH_INPUT_DEBUG");
debug = LLAMA_GRAPH_INPUT_DEBUG ? atoi(LLAMA_GRAPH_INPUT_DEBUG) : 0;
}
virtual ~llm_graph_input_i() = default;
virtual void set_input(const llama_ubatch * ubatch) = 0;
@@ -90,6 +95,9 @@ public:
GGML_UNUSED(params);
return false;
}
protected:
// env: LLAMA_GRAPH_INPUT_DEBUG
int debug = 0;
};
using llm_graph_input_ptr = std::unique_ptr<llm_graph_input_i>;
@@ -152,7 +160,7 @@ class llm_graph_input_pos_bucket_kv : public llm_graph_input_i {
public:
llm_graph_input_pos_bucket_kv(
const llama_hparams & hparams,
const llama_kv_cache_unified_context * mctx) : hparams(hparams), mctx(mctx) {}
const llama_kv_cache_context * mctx) : hparams(hparams), mctx(mctx) {}
virtual ~llm_graph_input_pos_bucket_kv() = default;
void set_input(const llama_ubatch * ubatch) override;
@@ -161,7 +169,7 @@ public:
const llama_hparams hparams;
const llama_kv_cache_unified_context * mctx;
const llama_kv_cache_context * mctx;
};
class llm_graph_input_out_ids : public llm_graph_input_i {
@@ -198,7 +206,7 @@ public:
class llm_graph_input_cls : public llm_graph_input_i {
public:
llm_graph_input_cls(const llama_cparams & cparams) : cparams(cparams) {}
llm_graph_input_cls(const llama_cparams & cparams, const llm_arch arch) : cparams(cparams), arch(arch) {}
virtual ~llm_graph_input_cls() = default;
void set_input(const llama_ubatch * ubatch) override;
@@ -206,6 +214,7 @@ public:
ggml_tensor * cls; // I32 [n_batch]
const llama_cparams cparams;
const llm_arch arch;
};
class llm_graph_input_rs : public llm_graph_input_i {
@@ -257,17 +266,17 @@ public:
const llama_cparams cparams;
};
class llm_graph_input_attn_kv_unified : public llm_graph_input_i {
class llm_graph_input_attn_kv : public llm_graph_input_i {
public:
llm_graph_input_attn_kv_unified(
llm_graph_input_attn_kv(
const llama_hparams & hparams,
const llama_cparams & cparams,
const llama_kv_cache_unified_context * mctx) :
const llama_kv_cache_context * mctx) :
hparams(hparams),
cparams(cparams),
mctx(mctx) {
}
~llm_graph_input_attn_kv_unified() = default;
~llm_graph_input_attn_kv() = default;
void set_input(const llama_ubatch * ubatch) override;
@@ -290,20 +299,20 @@ public:
const llama_hparams hparams;
const llama_cparams cparams;
const llama_kv_cache_unified_context * mctx;
const llama_kv_cache_context * mctx;
};
class llm_graph_input_attn_kv_unified_iswa : public llm_graph_input_i {
class llm_graph_input_attn_kv_iswa : public llm_graph_input_i {
public:
llm_graph_input_attn_kv_unified_iswa(
llm_graph_input_attn_kv_iswa(
const llama_hparams & hparams,
const llama_cparams & cparams,
const llama_kv_cache_unified_iswa_context * mctx) :
const llama_kv_cache_iswa_context * mctx) :
hparams(hparams),
cparams(cparams),
mctx(mctx) {
}
~llm_graph_input_attn_kv_unified_iswa() = default;
~llm_graph_input_attn_kv_iswa() = default;
void set_input(const llama_ubatch * ubatch) override;
@@ -330,7 +339,7 @@ public:
const llama_hparams hparams;
const llama_cparams cparams;
const llama_kv_cache_unified_iswa_context * mctx;
const llama_kv_cache_iswa_context * mctx;
};
class llm_graph_input_attn_cross : public llm_graph_input_i {
@@ -351,7 +360,7 @@ public:
class llm_graph_input_mem_hybrid : public llm_graph_input_i {
public:
llm_graph_input_mem_hybrid(
std::unique_ptr<llm_graph_input_attn_kv_unified> inp_attn,
std::unique_ptr<llm_graph_input_attn_kv> inp_attn,
std::unique_ptr<llm_graph_input_rs> inp_rs,
const llama_memory_hybrid_context * mctx) :
inp_attn(std::move(inp_attn)),
@@ -361,11 +370,11 @@ public:
void set_input(const llama_ubatch * ubatch) override;
std::unique_ptr<llm_graph_input_attn_kv_unified> inp_attn;
std::unique_ptr<llm_graph_input_rs> inp_rs;
std::unique_ptr<llm_graph_input_attn_kv> inp_attn;
std::unique_ptr<llm_graph_input_rs> inp_rs;
llm_graph_input_attn_kv_unified * get_attn() const { return inp_attn.get(); }
llm_graph_input_rs * get_recr() const { return inp_rs.get(); }
llm_graph_input_attn_kv * get_attn() const { return inp_attn.get(); }
llm_graph_input_rs * get_recr() const { return inp_rs.get(); }
const llama_memory_hybrid_context * mctx;
};
@@ -680,14 +689,15 @@ struct llm_graph_context {
//
ggml_tensor * build_attn_mha(
ggml_tensor * q, // [n_embd_head_q, n_head_q, n_tokens]
ggml_tensor * k, // [n_embd_head_k, n_head_k, n_tokens]
ggml_tensor * v, // [n_embd_head_v, n_head_v, n_tokens] (v_trans == false)
ggml_tensor * kq_b,
ggml_tensor * kq_mask,
ggml_tensor * sinks,
ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
float kq_scale) const;
ggml_tensor * q, // [n_embd_head_q, n_head_q, n_tokens]
ggml_tensor * k, // [n_embd_head_k, n_head_k, n_tokens]
ggml_tensor * v, // [n_embd_head_v, n_head_v, n_tokens] (v_trans == false)
ggml_tensor * kq_b,
ggml_tensor * kq_mask,
ggml_tensor * sinks, // [n_head_q]
ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
float kq_scale,
int il) const;
llm_graph_input_attn_no_cache * build_attn_inp_no_cache() const;
@@ -699,50 +709,39 @@ struct llm_graph_context {
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
ggml_tensor * kq_b,
ggml_tensor * sinks, // [n_head_q]
ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
float kq_scale,
int il) const;
llm_graph_input_attn_kv_unified * build_attn_inp_kv_unified() const;
llm_graph_input_attn_kv * build_attn_inp_kv() const;
ggml_tensor * build_attn(
llm_graph_input_attn_kv_unified * inp,
llm_graph_input_attn_kv * inp,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
ggml_tensor * kq_b,
ggml_tensor * sinks, // [n_head_q]
ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
float kq_scale,
int il) const;
llm_graph_input_attn_kv_unified_iswa * build_attn_inp_kv_unified_iswa() const;
llm_graph_input_attn_kv_iswa * build_attn_inp_kv_iswa() const;
// note: if k_cur or v_cur are not provided, they will not be stored in the memory
ggml_tensor * build_attn(
llm_graph_input_attn_kv_unified_iswa * inp,
llm_graph_input_attn_kv_iswa * inp,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] optional
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] optional
ggml_tensor * kq_b,
ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
float kq_scale,
int il) const;
// TODO: temporary to keep the diff small. after the code is public will refactor to simplify this
ggml_tensor * build_attn_with_sinks(
llm_graph_input_attn_kv_unified_iswa * inp,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] optional
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] optional
ggml_tensor * kq_b,
ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
ggml_tensor * sinks, // [n_head_q]
ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
float kq_scale,
int il) const;
@@ -756,6 +755,7 @@ struct llm_graph_context {
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
ggml_tensor * kq_b,
ggml_tensor * sinks, // [n_head_q]
ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
float kq_scale,
int il) const;
@@ -765,7 +765,7 @@ struct llm_graph_context {
//
// TODO: move this implementation to llama_memory_recurrent.
// this is analogous to llama_kv_cache_unified::cpy_k / cpy_v
// this is analogous to llama_kv_cache::cpy_k / cpy_v
// when moving, avoid passing `ggml_cgraph` - only pass `ggml_context`. would likely need to split the
// implementation in 2 separate methods. the goal is to avoid calling `ggml_build_forward_expand` in
// `llama_memory_recurrent`