quantize : handle user-defined pruning of whole layers (blocks) (#13037)

diff --git a/include/llama.h b/include/llama.h
index b04720bee59ef644922638d210632c7621d7edae..f4123d14ac1d8e31e298bff9017b92dd853e4d55 100644 (file)
--- a/include/llama.h
+++ b/include/llama.h
@@ -390,6 +390,7 @@ extern "C" {
         void * imatrix;                       // pointer to importance matrix data
         void * kv_overrides;                  // pointer to vector containing overrides
         void * tensor_types;                  // pointer to vector containing tensor types
+        void * prune_layers;                  // pointer to vector containing layer indices to prune
     } llama_model_quantize_params;
 
     typedef struct llama_logit_bias {

diff --git a/src/llama-quant.cpp b/src/llama-quant.cpp
index 8cf45732fd6d4817cdafd6fd3de6c2bea421fec4..43229e1938597a359f9f27ecccdc8a890982338d 100644 (file)
--- a/src/llama-quant.cpp
+++ b/src/llama-quant.cpp
@@ -1,5 +1,4 @@
 #include "llama-quant.h"
-
 #include "llama-impl.h"
 #include "llama-model.h"
 #include "llama-model-loader.h"
@@ -27,6 +26,56 @@ static void zeros(std::ofstream & file, size_t n) {
     }
 }
 
+static std::string remap_layer(const std::string & orig_name, const std::vector<int> & prune, std::map<int, std::string> & mapped, int & next_id) {
+    if (prune.empty()) {
+        return orig_name;
+    }
+
+    static const std::regex pattern(R"(blk\.(\d+)\.)");
+    if (std::smatch match; std::regex_search(orig_name, match, pattern)) {
+        const int blk = std::stoi(match[1]);
+        std::string new_name = orig_name;
+
+        if (mapped.count(blk)) {
+            // Already mapped, do nothing
+        } else if (std::find(prune.begin(), prune.end(), blk) != prune.end()) {
+            mapped[blk] = "";
+        } else if (blk < prune.front()) {
+            mapped[blk] = std::to_string(blk);
+            next_id = blk + 1;
+        } else {
+            mapped[blk] = std::to_string(next_id);
+            ++next_id;
+        }
+
+        return mapped[blk].empty() ? mapped[blk] : new_name.replace(match.position(1), match.length(1), mapped[blk]);
+    }
+
+    return orig_name;
+}
+
+static std::string remap_imatrix (const std::string & orig_name, const std::map<int, std::string> & mapped) {
+    if (mapped.empty()) {
+        return orig_name;
+    }
+
+    static const std::regex pattern(R"(blk\.(\d+)\.)");
+    if (std::smatch match; std::regex_search(orig_name, match, pattern)) {
+        const std::string blk(match[1]);
+        std::string new_name = orig_name;
+
+        for (const auto & p : mapped) {
+            if (p.second == blk) {
+                LLAMA_LOG_DEBUG("(blk.%d imatrix) ", p.first);
+                return new_name.replace(match.position(1), match.length(1), std::to_string(p.first));
+            }
+        }
+        GGML_ABORT("\n%s: imatrix mapping error for %s\n", __func__, orig_name.c_str());
+    }
+
+    return orig_name;
+}
+
 struct quantize_state_impl {
     const llama_model                 & model;
     const llama_model_quantize_params * params;
@@ -568,6 +617,11 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     const size_t align = GGUF_DEFAULT_ALIGNMENT;
     gguf_context_ptr ctx_out { gguf_init_empty() };
 
+    std::vector<int> prune_list = {};
+    if (params->prune_layers) {
+        prune_list = *static_cast<const std::vector<int> *>(params->prune_layers);
+    }
+
     // copy the KV pairs from the input file
     gguf_set_kv     (ctx_out.get(), ml.meta.get());
     gguf_set_val_u32(ctx_out.get(), "general.quantization_version", GGML_QNT_VERSION); // TODO: use LLM_KV
@@ -597,12 +651,32 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         }
     }
 
+    std::map<int, std::string> mapped;
+    int blk_id = 0;
+    int pruned_attention_w = 0;
+
     // make a list of weights
     std::vector<const llama_model_loader::llama_tensor_weight *> tensors;
     tensors.reserve(ml.weights_map.size());
     for (const auto & it : ml.weights_map) {
+        const std::string remapped_name(remap_layer(it.first, prune_list, mapped, blk_id));
+        if (remapped_name.empty()) {
+            if (it.first.find("attn_v.weight") != std::string::npos ||
+                it.first.find("attn_qkv.weight") != std::string::npos ||
+                it.first.find("attn_kv_b.weight") != std::string::npos) {
+                    pruned_attention_w++;
+            }
+            LLAMA_LOG_DEBUG("%s: pruning tensor %s\n", __func__, it.first.c_str());
+            continue;
+        } else if (remapped_name != it.first) {
+            ggml_set_name(it.second.tensor, remapped_name.c_str());
+            LLAMA_LOG_DEBUG("%s: tensor %s remapped to %s\n", __func__, it.first.c_str(), ggml_get_name(it.second.tensor));
+        }
         tensors.push_back(&it.second);
     }
+    if (!prune_list.empty()) {
+        gguf_set_val_u32(ctx_out.get(), ml.llm_kv(LLM_KV_BLOCK_COUNT).c_str(), blk_id);
+    }
 
     // keep_split requires that the weights are sorted by split index
     if (params->keep_split) {
@@ -640,7 +714,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         if (llama_model_has_encoder(&model)) {
             n_attn_layer *= 3;
         }
-        GGML_ASSERT((qs.n_attention_wv == n_attn_layer) && "n_attention_wv is unexpected");
+        GGML_ASSERT((qs.n_attention_wv == n_attn_layer - pruned_attention_w) && "n_attention_wv is unexpected");
     }
 
     size_t total_size_org = 0;
@@ -681,7 +755,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         for (size_t i = 0; i < ctx_outs.size(); ++i) {
             gguf_set_val_u16(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_NO).c_str(), i);
             gguf_set_val_u16(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_COUNT).c_str(), n_split);
-            gguf_set_val_i32(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str(), ml.n_tensors);
+            gguf_set_val_i32(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str(), (int32_t)tensors.size());
         }
     }
 
@@ -832,7 +906,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
 
             const float * imatrix = nullptr;
             if (imatrix_data) {
-                auto it = imatrix_data->find(tensor->name);
+                auto it = imatrix_data->find(remap_imatrix(tensor->name, mapped));
                 if (it == imatrix_data->end()) {
                     LLAMA_LOG_INFO("\n====== %s: did not find weights for %s\n", __func__, tensor->name);
                 } else {
@@ -947,6 +1021,7 @@ llama_model_quantize_params llama_model_quantize_default_params() {
         /*.imatrix                     =*/ nullptr,
         /*.kv_overrides                =*/ nullptr,
         /*.tensor_type                 =*/ nullptr,
+        /*.prune_layers                =*/ nullptr
     };
 
     return result;

diff --git a/tools/quantize/quantize.cpp b/tools/quantize/quantize.cpp
index 3f54af7c581588a6fc684fe8b47f9a6e3e512e6a..8acc765178846eca80271fe0d30fad87a2d87b44 100644 (file)
--- a/tools/quantize/quantize.cpp
+++ b/tools/quantize/quantize.cpp
@@ -107,13 +107,11 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
     return false;
 }
 
-// usage:
-//  ./llama-quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
-//
 [[noreturn]]
 static void usage(const char * executable) {
-    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] [--output-tensor-type]\n", executable);
-    printf("       [--token-embedding-type] [--tensor-type] [--keep-split] [--override-kv] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n");
+    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights]\n", executable);
+    printf("       [--exclude-weights] [--output-tensor-type] [--token-embedding-type] [--tensor-type] [--prune-layers] [--keep-split] [--override-kv]\n");
+    printf("       model-f32.gguf [model-quant.gguf] type [nthreads]\n\n");
     printf("  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
     printf("  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
     printf("  --pure: Disable k-quant mixtures and quantize all tensors to the same type\n");
@@ -124,6 +122,8 @@ static void usage(const char * executable) {
     printf("  --token-embedding-type ggml_type: use this ggml_type for the token embeddings tensor\n");
     printf("  --tensor-type TENSOR=TYPE: quantize this tensor to this ggml_type. example: --tensor-type attn_q=q8_0\n");
     printf("      Advanced option to selectively quantize tensors. May be specified multiple times.\n");
+    printf("  --prune-layers L0,L1,L2...comma-separated list of layer numbers to prune from the model\n");
+    printf("      Advanced option to remove all tensors from the given layers\n");
     printf("  --keep-split: will generate quantized model in the same shards as input\n");
     printf("  --override-kv KEY=TYPE:VALUE\n");
     printf("      Advanced option to override model metadata by key in the quantized model. May be specified multiple times.\n");
@@ -286,6 +286,32 @@ static bool parse_tensor_type(const char * data, std::vector<tensor_quantization
     return true;
 }
 
+static bool parse_layer_prune(const char * data, std::vector<int> & prune_layers) {
+    if (!data) {
+        printf("\n%s: no layer pruning ids provided\n\n", __func__);
+        return false;
+    }
+
+    const auto block_ids = string_split<std::string>(data, ',');
+    for (const auto & block_id : block_ids) {
+        int id;
+        try {
+            id = std::stoi(block_id);
+        } catch (...) {
+            id = -1;
+        }
+        if (id < 0) {
+            printf("\n%s: invalid layer id '%s'\n\n", __func__, block_id.c_str());
+            return false;
+        }
+        prune_layers.emplace_back(id);
+    }
+
+    sort(prune_layers.begin(), prune_layers.end());
+    prune_layers.erase(std::unique(prune_layers.begin(), prune_layers.end()), prune_layers.end());
+    return true;
+}
+
 int main(int argc, char ** argv) {
     if (argc < 3) {
         usage(argv[0]);
@@ -298,6 +324,7 @@ int main(int argc, char ** argv) {
     std::vector<std::string> included_weights, excluded_weights;
     std::vector<llama_model_kv_override> kv_overrides;
     std::vector<tensor_quantization> tensor_types;
+    std::vector<int> prune_layers;
 
     for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
         if (strcmp(argv[arg_idx], "--leave-output-tensor") == 0) {
@@ -324,6 +351,10 @@ int main(int argc, char ** argv) {
             if (arg_idx == argc-1 || !parse_tensor_type(argv[++arg_idx], tensor_types)) {
                 usage(argv[0]);
             }
+        } else if (strcmp(argv[arg_idx], "--prune-layers") == 0) {
+            if (arg_idx == argc-1 || !parse_layer_prune(argv[++arg_idx], prune_layers)) {
+                usage(argv[0]);
+            }
         } else if (strcmp(argv[arg_idx], "--override-kv") == 0) {
             if (arg_idx == argc-1 || !string_parse_kv_override(argv[++arg_idx], kv_overrides)) {
                 usage(argv[0]);
@@ -411,6 +442,9 @@ int main(int argc, char ** argv) {
     if (!tensor_types.empty()) {
         params.tensor_types = &tensor_types;
     }
+    if (!prune_layers.empty()) {
+        params.prune_layers = &prune_layers;
+    }
 
     llama_backend_init();