llama : hparams

ggml-ci

src/CMakeLists.txt

@@ -9,11 +9,12 @@ llama_add_compile_flags()
 add_library(llama
             ../include/llama.h
             llama.cpp
+            llama-adapter.cpp
             llama-arch.cpp
             llama-batch.cpp
             llama-chat.cpp
             llama-context.cpp
-            llama-adapter.cpp
+            llama-hparams.cpp
             llama-grammar.cpp
             llama-kv-cache.cpp
             llama-mmap.cpp

src/llama-adapter.h

@@ -5,9 +5,10 @@
 
 #include "llama-model.h" // TODO: need only hparams
 
-#include <vector>
-#include <map>
 #include <algorithm>
+#include <cassert>
+#include <map>
+#include <vector>
 
 //
 // llama_adapter_vec

src/llama-chat.h

@@ -2,6 +2,7 @@
 
 #include <string>
 #include <vector>
+#include <cstdint>
 
 enum llm_chat_template {
     LLM_CHAT_TEMPLATE_CHATML,

src/llama-hparams.cpp

@@ -0,0 +1,71 @@
+#include "llama-hparams.h"
+
+#include "ggml.h"
+
+uint32_t llama_hparams::n_head(uint32_t il) const {
+    if (il < n_layer) {
+        return n_head_arr[il];
+    }
+
+    GGML_ABORT("fatal error");
+}
+
+uint32_t llama_hparams::n_head_kv(uint32_t il) const {
+    if (il < n_layer) {
+        return n_head_kv_arr[il];
+    }
+
+    GGML_ABORT("fatal error");
+}
+
+uint32_t llama_hparams::n_ff(uint32_t il) const {
+    if (il < n_layer) {
+        return n_ff_arr[il];
+    }
+
+    GGML_ABORT("fatal error");
+}
+
+uint32_t llama_hparams::n_gqa(uint32_t il) const {
+    const uint32_t n_head    = this->n_head(il);
+    const uint32_t n_head_kv = this->n_head_kv(il);
+
+    if (n_head_kv == 0) {
+        return 0;
+    }
+
+    return n_head/n_head_kv;
+}
+
+uint32_t llama_hparams::n_embd_k_gqa(uint32_t il) const {
+    const uint32_t n_head_kv = this->n_head_kv(il);
+
+    return n_embd_head_k * n_head_kv;
+}
+
+uint32_t llama_hparams::n_embd_v_gqa(uint32_t il) const {
+    const uint32_t n_head_kv = this->n_head_kv(il);
+
+    return n_embd_head_v * n_head_kv;
+}
+
+uint32_t llama_hparams::n_embd_k_s() const {
+    if (wkv_head_size != 0) {
+        // for RWKV models
+        return 2 * n_embd;
+    }
+
+    // TODO: maybe support other convolution strides than 1
+    // NOTE: since the first column of the conv_state is shifted out each time, it's not actually needed
+    return (ssm_d_conv > 0 ? ssm_d_conv - 1 : 0) * ssm_d_inner;
+}
+
+uint32_t llama_hparams::n_embd_v_s() const {
+    if (wkv_head_size != 0) {
+        // corresponds to RWKV's wkv_states size
+        return n_embd * wkv_head_size;
+    }
+
+    // corresponds to Mamba's ssm_states size
+    return ssm_d_state * ssm_d_inner;
+}

src/llama-hparams.h

@@ -0,0 +1,131 @@
+#pragma once
+
+#include "llama.h"
+
+#include <array>
+
+// bump if necessary
+#define LLAMA_MAX_LAYERS  512
+#define LLAMA_MAX_EXPERTS 160  // DeepSeekV2
+
+struct llama_hparams_posnet {
+    uint32_t n_embd;
+    uint32_t n_layer;
+};
+
+struct llama_hparams_convnext {
+    uint32_t n_embd;
+    uint32_t n_layer;
+};
+
+struct llama_hparams {
+    bool vocab_only;
+    bool rope_finetuned;
+    bool use_par_res;
+    bool swin_norm;
+
+    uint32_t n_vocab = 0;
+    uint32_t n_ctx_train; // context size the model was trained on
+    uint32_t n_embd;
+    uint32_t n_embd_features = 0;
+    uint32_t n_layer;
+    uint32_t n_rot;
+    uint32_t n_swa = 0; // sliding window attention (SWA)
+    uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
+    uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head
+    uint32_t n_expert = 0;
+    uint32_t n_expert_used = 0;
+    uint32_t n_vocab_type = 0; // for BERT-style token types
+    uint32_t n_rel_attn_bkts = 0;
+
+    // for WavTokenizer
+    struct llama_hparams_posnet   posnet;
+    struct llama_hparams_convnext convnext;
+
+    std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_arr;
+    std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_kv_arr;
+    std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
+
+    uint32_t n_layer_dense_lead = 0;
+    uint32_t n_lora_q           = 0;
+    uint32_t n_lora_kv          = 0;
+    uint32_t n_ff_exp           = 0;
+    uint32_t n_ff_shexp         = 0;
+    uint32_t n_expert_shared    = 0;
+    uint32_t n_norm_groups      = 0;
+
+    float expert_weights_scale = 0.0;
+
+    float f_norm_eps;
+    float f_norm_rms_eps;
+    float f_norm_group_eps;
+
+    float f_attn_logit_softcapping  = 50.0f;
+    float f_final_logit_softcapping = 30.0f;
+
+    // for RWKV
+    uint32_t rescale_every_n_layers = 0;
+    uint32_t time_mix_extra_dim     = 0;
+    uint32_t time_decay_extra_dim   = 0;
+    uint32_t wkv_head_size          = 0;
+
+    float     rope_attn_factor = 1.0f;
+    float     rope_freq_base_train;
+    float     rope_freq_scale_train;
+    uint32_t  n_ctx_orig_yarn;
+    float     rope_yarn_log_mul;
+    int       rope_sections[4]; // TODO: actually this should be std::array (I was wrong)
+
+    // for State Space Models
+    uint32_t ssm_d_conv  = 0;
+    uint32_t ssm_d_inner = 0;
+    uint32_t ssm_d_state = 0;
+    uint32_t ssm_dt_rank = 0;
+
+    bool ssm_dt_b_c_rms = false;
+
+    float f_clamp_kqv      = 0.0f;
+    float f_max_alibi_bias = 0.0f;
+    float f_logit_scale    = 0.0f;
+
+    // Additional scale factors (Granite/Granite MoE)
+    float f_residual_scale  = 0.0f;
+    float f_embedding_scale = 0.0f;
+    float f_attention_scale = 0.0f;
+
+    bool causal_attn   = true;
+    bool use_alibi     = false;
+    bool attn_soft_cap = false;
+
+    // needed by encoder-decoder models (e.g. T5, FLAN-T5)
+    // ref: https://github.com/ggerganov/llama.cpp/pull/8141
+    llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
+
+    enum llama_pooling_type      pooling_type            = LLAMA_POOLING_TYPE_NONE;
+    enum llama_rope_type         rope_type               = LLAMA_ROPE_TYPE_NONE;
+    enum llama_rope_scaling_type rope_scaling_type_train = LLAMA_ROPE_SCALING_TYPE_NONE;
+
+    uint32_t n_head(uint32_t il = 0) const;
+
+    uint32_t n_head_kv(uint32_t il = 0) const;
+
+    uint32_t n_ff(uint32_t il = 0) const;
+
+    uint32_t n_gqa(uint32_t il = 0) const;
+
+    // dimension of key embeddings across all k-v heads
+    uint32_t n_embd_k_gqa(uint32_t il = 0) const;
+
+    // dimension of value embeddings across all k-v heads
+    uint32_t n_embd_v_gqa(uint32_t il = 0) const;
+
+    // dimension of the rolling state embeddings
+    // corresponds to Mamba's conv_states size or RWKV's token_shift states size
+    uint32_t n_embd_k_s() const;
+
+    // dimension of the recurrent state embeddings
+    uint32_t n_embd_v_s() const;
+};
+
+static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
+

src/llama-model.cpp

@@ -2,6 +2,8 @@
 
 #include "llama-impl.h"
 
+#include <cassert>
+
 const char * llm_type_name(llm_type type) {
     switch (type) {
         case MODEL_14M:           return "14M";

src/llama-model.h

@@ -2,18 +2,13 @@
 
 #include "llama.h"
 #include "llama-arch.h"
+#include "llama-hparams.h"
 #include "llama-vocab.h"
 #include "llama-mmap.h"
 
 #include "ggml-cpp.h"
 
-#include <array>
 #include <vector>
-#include <cassert>
-
-// bump if necessary
-#define LLAMA_MAX_LAYERS  512
-#define LLAMA_MAX_EXPERTS 160  // DeepSeekV2
 
 // available models
 // TODO: this enum does not follow the enum naming convention
@@ -82,175 +77,6 @@ enum llm_type {
     MODEL_27B,
 };
 
-struct llama_hparams_posnet {
-    uint32_t n_embd;
-    uint32_t n_layer;
-};
-
-struct llama_hparams_convnext {
-    uint32_t n_embd;
-    uint32_t n_layer;
-};
-
-struct llama_hparams {
-    bool vocab_only;
-    bool rope_finetuned;
-    bool use_par_res;
-    bool swin_norm;
-
-    uint32_t n_vocab = 0;
-    uint32_t n_ctx_train; // context size the model was trained on
-    uint32_t n_embd;
-    uint32_t n_embd_features = 0;
-    uint32_t n_layer;
-    uint32_t n_rot;
-    uint32_t n_swa = 0; // sliding window attention (SWA)
-    uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
-    uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head
-    uint32_t n_expert = 0;
-    uint32_t n_expert_used = 0;
-    uint32_t n_vocab_type = 0; // for BERT-style token types
-    uint32_t n_rel_attn_bkts = 0;
-
-    // for WavTokenizer
-    struct llama_hparams_posnet   posnet;
-    struct llama_hparams_convnext convnext;
-
-    std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_arr;
-    std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_kv_arr;
-    std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
-
-    uint32_t n_layer_dense_lead = 0;
-    uint32_t n_lora_q           = 0;
-    uint32_t n_lora_kv          = 0;
-    uint32_t n_ff_exp           = 0;
-    uint32_t n_ff_shexp         = 0;
-    uint32_t n_expert_shared    = 0;
-    uint32_t n_norm_groups      = 0;
-
-    float expert_weights_scale = 0.0;
-
-    float f_norm_eps;
-    float f_norm_rms_eps;
-    float f_norm_group_eps;
-
-    float f_attn_logit_softcapping  = 50.0f;
-    float f_final_logit_softcapping = 30.0f;
-
-    // for RWKV
-    uint32_t rescale_every_n_layers = 0;
-    uint32_t time_mix_extra_dim     = 0;
-    uint32_t time_decay_extra_dim   = 0;
-    uint32_t wkv_head_size          = 0;
-
-    float     rope_attn_factor = 1.0f;
-    float     rope_freq_base_train;
-    float     rope_freq_scale_train;
-    uint32_t  n_ctx_orig_yarn;
-    float     rope_yarn_log_mul;
-    int       rope_sections[4];
-
-    // for State Space Models
-    uint32_t ssm_d_conv  = 0;
-    uint32_t ssm_d_inner = 0;
-    uint32_t ssm_d_state = 0;
-    uint32_t ssm_dt_rank = 0;
-
-    bool ssm_dt_b_c_rms = false;
-
-    float f_clamp_kqv      = 0.0f;
-    float f_max_alibi_bias = 0.0f;
-    float f_logit_scale    = 0.0f;
-
-    // Additional scale factors (Granite/Granite MoE)
-    float f_residual_scale  = 0.0f;
-    float f_embedding_scale = 0.0f;
-    float f_attention_scale = 0.0f;
-
-    bool causal_attn   = true;
-    bool use_alibi     = false;
-    bool attn_soft_cap = false;
-
-    // needed by encoder-decoder models (e.g. T5, FLAN-T5)
-    // ref: https://github.com/ggerganov/llama.cpp/pull/8141
-    llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
-
-    enum llama_pooling_type      pooling_type            = LLAMA_POOLING_TYPE_NONE;
-    enum llama_rope_type         rope_type               = LLAMA_ROPE_TYPE_NONE;
-    enum llama_rope_scaling_type rope_scaling_type_train = LLAMA_ROPE_SCALING_TYPE_NONE;
-
-    uint32_t n_head(uint32_t il = 0) const {
-        if (il < n_layer) {
-            return n_head_arr[il];
-        }
-
-        GGML_ABORT("fatal error");
-    }
-
-    uint32_t n_head_kv(uint32_t il = 0) const {
-        if (il < n_layer) {
-            return n_head_kv_arr[il];
-        }
-
-        GGML_ABORT("fatal error");
-    }
-
-    uint32_t n_ff(uint32_t il = 0) const {
-        if (il < n_layer) {
-            return n_ff_arr[il];
-        }
-
-        GGML_ABORT("fatal error");
-    }
-
-    uint32_t n_gqa(uint32_t il = 0) const {
-        const uint32_t n_head    = this->n_head(il);
-        const uint32_t n_head_kv = this->n_head_kv(il);
-
-        if (n_head_kv == 0) {
-            return 0;
-        }
-
-        return n_head/n_head_kv;
-    }
-
-    uint32_t n_embd_k_gqa(uint32_t il = 0) const { // dimension of key embeddings across all k-v heads
-        const uint32_t n_head_kv = this->n_head_kv(il);
-
-        return n_embd_head_k * n_head_kv;
-    }
-
-    uint32_t n_embd_v_gqa(uint32_t il = 0) const { // dimension of value embeddings across all k-v heads
-        const uint32_t n_head_kv = this->n_head_kv(il);
-
-        return n_embd_head_v * n_head_kv;
-    }
-
-    uint32_t n_embd_k_s() const { // dimension of the rolling state embeddings
-        // corresponds to Mamba's conv_states size or RWKV's token_shift states size
-        if (wkv_head_size != 0) {
-            // for RWKV models
-            return 2 * n_embd;
-        }
-
-        // TODO: maybe support other convolution strides than 1
-        // NOTE: since the first column of the conv_state is shifted out each time, it's not actually needed
-        return (ssm_d_conv > 0 ? ssm_d_conv - 1 : 0) * ssm_d_inner;
-    }
-
-    uint32_t n_embd_v_s() const { // dimension of the recurrent state embeddings
-        if (wkv_head_size != 0) {
-            // corresponds to RWKV's wkv_states size
-            return n_embd * wkv_head_size;
-        }
-
-        // corresponds to Mamba's ssm_states size
-        return ssm_d_state * ssm_d_inner;
-    }
-};
-
-static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
-
 struct llama_layer_posnet {
     // resnet
     struct ggml_tensor * norm1   = nullptr;
@@ -522,6 +348,7 @@ struct llama_model {
     llama_mmaps mappings;
 
     // objects representing data potentially being locked in memory
+    // TODO: should these be part of llama_context instead?
     llama_mlocks mlock_bufs;
     llama_mlocks mlock_mmaps;