269 lines
10 KiB
C
269 lines
10 KiB
C
/* SPDX-License-Identifier: AGPL-3.0-or-later */
|
|
/* Copyright (C) 2026 David (Corey) Herrera */
|
|
/*
|
|
* test_hybridmem_retrieval.c — A/B test of HybridMem retrieval with real inference
|
|
*
|
|
* Feeds: [fact] [filler from file] [question about fact]
|
|
* Run A: small cache, NO disk backup → fact is evicted and lost
|
|
* Run B: small cache, WITH disk backup → fact is evicted but retrievable
|
|
*
|
|
* Both runs use identical cache size so the only variable is disk retrieval.
|
|
*
|
|
* Usage:
|
|
* ./examples/test_hybridmem_retrieval \
|
|
* --model ../models/base/RWKV-X-0.2B-64k-Base.safetensors \
|
|
* --vocab rwkvx/data/rwkv_vocab_v20230424.txt \
|
|
* --filler docs/architecture/rwkv-x-deep-dive.md
|
|
*/
|
|
|
|
#include "rwkvx/rwkvx.h"
|
|
#include "tensor/tensor.h"
|
|
|
|
#include <stdio.h>
|
|
#include <stdlib.h>
|
|
#include <string.h>
|
|
#include <time.h>
|
|
#include <sys/stat.h>
|
|
|
|
/* Read entire file into malloc'd string. Returns NULL on error. */
|
|
static char* read_file(const char *path) {
|
|
FILE *fp = fopen(path, "r");
|
|
if (!fp) return NULL;
|
|
fseek(fp, 0, SEEK_END);
|
|
long len = ftell(fp);
|
|
fseek(fp, 0, SEEK_SET);
|
|
char *buf = (char*)malloc(len + 1);
|
|
size_t read = fread(buf, 1, len, fp);
|
|
buf[read] = '\0';
|
|
fclose(fp);
|
|
return buf;
|
|
}
|
|
|
|
/* Generate tokens from a state that has already been prefilled.
|
|
* Returns malloc'd string of generated text. */
|
|
static char* generate(rx_model *model, rx_tokenizer *tok, rx_state *state,
|
|
tm_tensor *logits, int max_tokens, rx_config cfg) {
|
|
rx_sampling_params params = rx_sampling_defaults();
|
|
params.temperature = 0.0f; /* greedy for reproducibility */
|
|
|
|
uint32_t *gen_tokens = (uint32_t*)calloc(max_tokens, sizeof(uint32_t));
|
|
int n_gen = 0;
|
|
tm_tensor *cur_logits = logits;
|
|
|
|
for (int i = 0; i < max_tokens; i++) {
|
|
if (!cur_logits) break;
|
|
uint32_t token = rx_sample(cur_logits, ¶ms, NULL, cfg.vocab_size);
|
|
if (i > 0) tm_free(cur_logits);
|
|
if (token == 0) { cur_logits = NULL; break; } /* end-of-text */
|
|
gen_tokens[n_gen++] = token;
|
|
cur_logits = rx_forward(model, token, state);
|
|
}
|
|
if (cur_logits && cur_logits != logits) tm_free(cur_logits);
|
|
|
|
char *text = rx_tokenizer_decode(tok, gen_tokens, n_gen);
|
|
free(gen_tokens);
|
|
return text;
|
|
}
|
|
|
|
/* Recursive rm -rf for cleanup */
|
|
static void rmdir_recursive(const char *path) {
|
|
char cmd[512];
|
|
snprintf(cmd, sizeof(cmd), "rm -rf %s", path);
|
|
system(cmd);
|
|
}
|
|
|
|
int main(int argc, char **argv) {
|
|
const char *model_path = NULL;
|
|
const char *vocab_path = NULL;
|
|
const char *filler_path = NULL;
|
|
int cache_size = 8;
|
|
int hm_top_k = 5;
|
|
int max_filler_chars = 2000;
|
|
int max_tokens = 32;
|
|
int quantize = 0;
|
|
|
|
for (int i = 1; i < argc; i++) {
|
|
if (strcmp(argv[i], "--model") == 0 && i + 1 < argc)
|
|
model_path = argv[++i];
|
|
else if (strcmp(argv[i], "--vocab") == 0 && i + 1 < argc)
|
|
vocab_path = argv[++i];
|
|
else if (strcmp(argv[i], "--filler") == 0 && i + 1 < argc)
|
|
filler_path = argv[++i];
|
|
else if (strcmp(argv[i], "--cache-size") == 0 && i + 1 < argc)
|
|
cache_size = atoi(argv[++i]);
|
|
else if (strcmp(argv[i], "--hm-top-k") == 0 && i + 1 < argc)
|
|
hm_top_k = atoi(argv[++i]);
|
|
else if (strcmp(argv[i], "--max-filler") == 0 && i + 1 < argc)
|
|
max_filler_chars = atoi(argv[++i]);
|
|
else if (strcmp(argv[i], "--max-tokens") == 0 && i + 1 < argc)
|
|
max_tokens = atoi(argv[++i]);
|
|
else if (strcmp(argv[i], "--quantize") == 0)
|
|
quantize = 1;
|
|
else if (strcmp(argv[i], "--help") == 0) {
|
|
fprintf(stderr,
|
|
"Usage: %s --model PATH --vocab PATH --filler PATH [options]\n"
|
|
" --cache-size N KV cache slots per attn layer (default: 8)\n"
|
|
" --hm-top-k N Retrieval count (default: 5)\n"
|
|
" --max-filler N Max chars of filler text (default: 2000)\n"
|
|
" --max-tokens N Tokens to generate (default: 32)\n"
|
|
" --quantize Quantize weights to INT8\n",
|
|
argv[0]);
|
|
return 0;
|
|
}
|
|
}
|
|
|
|
if (!model_path || !vocab_path || !filler_path) {
|
|
fprintf(stderr, "Error: --model, --vocab, and --filler are required\n");
|
|
return 1;
|
|
}
|
|
|
|
srand(42);
|
|
|
|
/* Load shared resources */
|
|
fprintf(stderr, "Loading tokenizer...\n");
|
|
rx_tokenizer *tok = rx_tokenizer_new(vocab_path);
|
|
if (!tok) { fprintf(stderr, "Failed to load tokenizer\n"); return 1; }
|
|
|
|
fprintf(stderr, "Loading model%s...\n", quantize ? " (INT8 quantized)" : "");
|
|
rx_model *model = quantize ? rx_model_load_q8(model_path)
|
|
: rx_model_load(model_path);
|
|
if (!model) { fprintf(stderr, "Failed to load model\n"); return 1; }
|
|
rx_config cfg = rx_model_config(model);
|
|
|
|
/* Build prompt: fact + filler + question */
|
|
const char *fact = "The secret project codename is AURORA-7 and the lead researcher is Dr. Kimura.";
|
|
const char *question = "\n\nQuestion: What is the secret project codename and who leads it?\nAnswer:";
|
|
|
|
char *filler_raw = read_file(filler_path);
|
|
if (!filler_raw) {
|
|
fprintf(stderr, "Failed to read filler file: %s\n", filler_path);
|
|
return 1;
|
|
}
|
|
|
|
/* Truncate filler */
|
|
int filler_len = (int)strlen(filler_raw);
|
|
if (filler_len > max_filler_chars) {
|
|
filler_raw[max_filler_chars] = '\0';
|
|
filler_len = max_filler_chars;
|
|
}
|
|
|
|
/* Assemble full prompt */
|
|
int prompt_len = (int)(strlen(fact) + 2 + filler_len + strlen(question) + 1);
|
|
char *prompt = (char*)malloc(prompt_len);
|
|
snprintf(prompt, prompt_len, "%s\n\n%s%s", fact, filler_raw, question);
|
|
free(filler_raw);
|
|
|
|
/* Tokenize */
|
|
uint32_t *tokens = (uint32_t*)malloc(32768 * sizeof(uint32_t));
|
|
int n_tokens = rx_tokenizer_encode(tok, prompt, tokens, 32768);
|
|
|
|
fprintf(stderr, "\n==========================================\n");
|
|
fprintf(stderr, "HybridMem A/B Retrieval Test\n");
|
|
fprintf(stderr, "==========================================\n");
|
|
fprintf(stderr, "Prompt tokens: %d\n", n_tokens);
|
|
fprintf(stderr, "Cache size: %d (per attn layer)\n", cache_size);
|
|
fprintf(stderr, "HM top-k: %d\n", hm_top_k);
|
|
fprintf(stderr, "Generate: %d tokens (greedy)\n", max_tokens);
|
|
fprintf(stderr, "Fact: \"%s\"\n", fact);
|
|
fprintf(stderr, "Question: \"%s\"\n", question);
|
|
fprintf(stderr, "\n");
|
|
|
|
/* ============================================================
|
|
* Run A: Small cache, NO HybridMem (evicted entries are lost)
|
|
* ============================================================ */
|
|
fprintf(stderr, "--- Run A: cache_size=%d, NO HybridMem ---\n", cache_size);
|
|
{
|
|
rx_hybridmem_config hm_cfg = rx_hybridmem_config_defaults();
|
|
hm_cfg.store_dir = NULL; /* no disk backup */
|
|
hm_cfg.max_cache_len = cache_size;
|
|
|
|
rx_state *state_a = rx_state_new_with_hybridmem(model, &hm_cfg);
|
|
if (!state_a) { fprintf(stderr, "Failed to create state A\n"); return 1; }
|
|
|
|
struct timespec t0, t1;
|
|
clock_gettime(CLOCK_MONOTONIC, &t0);
|
|
|
|
tm_tensor *logits_a = NULL;
|
|
for (int i = 0; i < n_tokens; i++) {
|
|
if (logits_a) tm_free(logits_a);
|
|
logits_a = rx_forward(model, tokens[i], state_a);
|
|
}
|
|
|
|
clock_gettime(CLOCK_MONOTONIC, &t1);
|
|
double ms = (t1.tv_sec - t0.tv_sec) * 1000.0 + (t1.tv_nsec - t0.tv_nsec) / 1e6;
|
|
fprintf(stderr, "Prefill: %d tokens in %.1fms (%.1f tok/s)\n",
|
|
n_tokens, ms, n_tokens * 1000.0 / ms);
|
|
|
|
char *output_a = generate(model, tok, state_a, logits_a, max_tokens, cfg);
|
|
tm_free(logits_a);
|
|
rx_state_free(state_a);
|
|
|
|
fprintf(stderr, "Output A: %s\n\n", output_a ? output_a : "(null)");
|
|
free(output_a);
|
|
}
|
|
|
|
/* ============================================================
|
|
* Run B: Small cache, WITH HybridMem (evicted entries retrievable)
|
|
* ============================================================ */
|
|
const char *hm_dir = "/tmp/hm_ab_test";
|
|
rmdir_recursive(hm_dir);
|
|
|
|
fprintf(stderr, "--- Run B: cache_size=%d, WITH HybridMem (top_k=%d) ---\n",
|
|
cache_size, hm_top_k);
|
|
{
|
|
rx_hybridmem_config hm_cfg = rx_hybridmem_config_defaults();
|
|
hm_cfg.store_dir = hm_dir;
|
|
hm_cfg.max_cache_len = cache_size;
|
|
hm_cfg.top_k = hm_top_k;
|
|
|
|
rx_state *state_b = rx_state_new_with_hybridmem(model, &hm_cfg);
|
|
if (!state_b) { fprintf(stderr, "Failed to create state B\n"); return 1; }
|
|
|
|
struct timespec t0, t1;
|
|
clock_gettime(CLOCK_MONOTONIC, &t0);
|
|
|
|
tm_tensor *logits_b = NULL;
|
|
for (int i = 0; i < n_tokens; i++) {
|
|
if (logits_b) tm_free(logits_b);
|
|
logits_b = rx_forward(model, tokens[i], state_b);
|
|
}
|
|
|
|
clock_gettime(CLOCK_MONOTONIC, &t1);
|
|
double ms = (t1.tv_sec - t0.tv_sec) * 1000.0 + (t1.tv_nsec - t0.tv_nsec) / 1e6;
|
|
fprintf(stderr, "Prefill: %d tokens in %.1fms (%.1f tok/s)\n",
|
|
n_tokens, ms, n_tokens * 1000.0 / ms);
|
|
|
|
char *output_b = generate(model, tok, state_b, logits_b, max_tokens, cfg);
|
|
tm_free(logits_b);
|
|
rx_state_flush(state_b);
|
|
rx_state_free(state_b);
|
|
|
|
fprintf(stderr, "Output B: %s\n\n", output_b ? output_b : "(null)");
|
|
|
|
/* Check disk artifacts exist */
|
|
struct stat sb;
|
|
char shard_path[512];
|
|
snprintf(shard_path, sizeof(shard_path), "%s/shard_0000.safetensors", hm_dir);
|
|
if (stat(shard_path, &sb) == 0) {
|
|
fprintf(stderr, "HybridMem shard: %s (%lld bytes)\n", shard_path, (long long)sb.st_size);
|
|
} else {
|
|
fprintf(stderr, "WARNING: No shard file found at %s\n", shard_path);
|
|
}
|
|
free(output_b);
|
|
}
|
|
|
|
/* Cleanup */
|
|
rmdir_recursive(hm_dir);
|
|
free(prompt);
|
|
free(tokens);
|
|
rx_model_free(model);
|
|
rx_tokenizer_free(tok);
|
|
|
|
fprintf(stderr, "\n==========================================\n");
|
|
fprintf(stderr, "Compare outputs above. If HybridMem retrieval works,\n");
|
|
fprintf(stderr, "Run B should reference AURORA-7 / Dr. Kimura more than Run A.\n");
|
|
fprintf(stderr, "==========================================\n");
|
|
|
|
return 0;
|
|
}
|