Steering

examples/common.cpp

@@ -344,6 +344,30 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.input_suffix = argv[i];
+        } else if (arg == "--steering-add") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.steering_add = argv[i];
+        } else if (arg == "--steering-sub") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.steering_sub = argv[i];
+        } else if (arg == "--steering-mul") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.steering_mul = std::stof(argv[i]);
+        } else if (arg == "--steering-lyr") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.steering_lyr = std::stoi(argv[i]);
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             gpt_print_usage(argc, argv, default_params);

examples/common.h

@@ -72,6 +72,11 @@ struct gpt_params {
     bool use_mlock         = false; // use mlock to keep model in memory
     bool mem_test          = false; // compute maximum memory usage
     bool verbose_prompt    = false; // print prompt tokens before generation
+
+    std::string steering_add = "";
+    std::string steering_sub = "";
+    float       steering_mul = 1.0f;
+    int         steering_lyr = 20;
 };
 
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params);

examples/main/main.cpp

@@ -136,6 +136,28 @@ int main(int argc, char ** argv) {
         return 0;
     }
 
+    if (params.steering_add.size() || params.steering_sub.size())
+    {
+        auto steering_add_tokens = ::llama_tokenize(ctx, params.steering_add, true);
+        auto steering_sub_tokens = ::llama_tokenize(ctx, params.steering_sub, true);
+
+        if (steering_add_tokens.size() != steering_sub_tokens.size()) {
+            llama_token space;
+            llama_tokenize(ctx, " ", &space, 1, 0);
+
+            while (steering_add_tokens.size() < steering_sub_tokens.size()) steering_add_tokens.push_back(space);
+            while (steering_sub_tokens.size() < steering_add_tokens.size()) steering_sub_tokens.push_back(space);
+        }
+
+        llama_set_steering_write(ctx, params.steering_lyr, params.steering_mul/2);
+        llama_eval(ctx, steering_add_tokens.data(), std::min((int)steering_add_tokens.size(), params.n_ctx), 0, params.n_threads);
+
+        llama_set_steering_write(ctx, params.steering_lyr, -params.steering_mul/2);
+        llama_eval(ctx, steering_sub_tokens.data(), std::min((int)steering_sub_tokens.size(), params.n_ctx), 0, params.n_threads);
+
+        llama_set_steering_read(ctx, params.steering_lyr, 1);
+    }
+
     // Add a space in front of the first character to match OG llama tokenizer behavior
     params.prompt.insert(0, 1, ' ');
 

llama.cpp

@@ -229,6 +229,15 @@ struct llama_context {
     // input embedding (1-dimensional array: [n_embd])
     std::vector<float> embedding;
 
+    std::vector<float> steering_vector; // [n_ctx, n_embd]
+    int                steering_layer = 0;
+    int                steering_mode = 0;
+    float              steering_mul = 0.0f;
+
+    #define STEERING_OFF   0
+    #define STEERING_WRITE 2
+    #define STEERING_READ  3
+
     // memory buffers used to evaluate the model
     // TODO: move in llama_state
     llama_ctx_buffer buf_compute;
@@ -269,6 +278,17 @@ struct llama_context {
     }
 };
 
+void llama_set_steering_write(struct llama_context * ctx, int layer, float mul) {
+    ctx->steering_mode = STEERING_WRITE;
+    ctx->steering_mul = mul;
+    ctx->steering_layer = layer;
+}
+void llama_set_steering_read(struct llama_context * ctx, int layer, float mul) {
+    ctx->steering_mode = STEERING_READ;
+    ctx->steering_mul = mul;
+    ctx->steering_layer = layer;
+}
+
 template <typename T>
 static T checked_mul(T a, T b) {
     T ret = a * b;
@@ -1141,6 +1161,12 @@ static bool llama_eval_internal(
     ggml_set_name(embd, "embd");
     memcpy(embd->data, tokens, N*ggml_element_size(embd));
 
+    struct ggml_tensor * steer;
+    if (lctx.steering_mode != STEERING_OFF) {
+        steer = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_ctx, n_embd);
+        memcpy(steer->data, lctx.steering_vector.data(), ggml_nbytes(steer));
+    }
+
     struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.tok_embeddings, embd);
 
     for (int il = 0; il < n_layer; ++il) {
@@ -1150,6 +1176,18 @@ static bool llama_eval_internal(
 
         lctx.use_buf(ctx0, 0);
 
+        if (lctx.steering_mode != STEERING_OFF && il == lctx.steering_layer) {
+            steer->data = lctx.steering_vector.data();
+
+            struct ggml_tensor * src = ggml_scale(ctx0, inpL, ggml_new_f32(ctx0, lctx.steering_mul));
+            struct ggml_tensor * dst = ggml_view_2d(ctx0, steer, n_embd, N, n_embd * sizeof(float), n_past * n_embd * sizeof(float));
+            if (lctx.steering_mode == STEERING_WRITE) {
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, ggml_add(ctx0, src, dst), dst));
+            } else {
+                inpL = src;
+            }
+        }
+
         // norm
         {
             cur = ggml_rms_norm(ctx0, inpL);
@@ -1363,6 +1401,12 @@ static bool llama_eval_internal(
         memcpy(embedding_out.data(), (float *) ggml_get_data(embeddings) + (n_embd*(N - 1)), sizeof(float)*n_embd);
     }
 
+
+    if (lctx.steering_mode == STEERING_WRITE) {
+        memcpy(lctx.steering_vector.data(), steer->data, ggml_nbytes(steer));
+    }
+
+
     if (mem_per_token == 0) {
         mem_per_token = ggml_used_mem(ctx0)/N;
     }
@@ -2184,6 +2228,8 @@ struct llama_context * llama_init_from_file(
 
         ctx->buf_scratch[0].resize(MEM_REQ_SCRATCH0().at(ctx->model.type));
         ctx->buf_scratch[1].resize(MEM_REQ_SCRATCH1().at(ctx->model.type));
+
+        ctx->steering_vector.resize(hparams.n_ctx * hparams.n_embd);
     }
 
     return ctx;

llama.h

@@ -191,6 +191,9 @@ extern "C" {
     LLAMA_API llama_token llama_token_eos();
     LLAMA_API llama_token llama_token_nl();
 
+    LLAMA_API void llama_set_steering_write(struct llama_context * ctx, int layer, float mul);
+    LLAMA_API void llama_set_steering_read(struct llama_context * ctx, int layer, float mul);
+
     // Sampling functions
 
     /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.