Add quant.h — single-header LLM inference library (15K LOC)

stb-style single-header C library for LLM inference.
One file. Six functions. Zero dependencies beyond libc/pthreads.

Usage:
  #define QUANT_IMPLEMENTATION
  #include "quant.h"

  quant_model* m = quant_load("model.gguf");
  quant_ctx*   c = quant_new(m, NULL);
  quant_generate(c, "Hello!", print_token, NULL);
  quant_free_ctx(c);
  quant_free_model(m);

Build: cc app.c -o app -lm -lpthread

Features included:
- GGUF model loading (all K-quant types, IQ2, BF16, F16)
- BPE tokenizer (from GGUF metadata)
- Transformer forward pass (Llama, Qwen3.5, Gemma 3/4)
- Uniform 4-bit/3-bit/2-bit KV cache compression
- Delta KV compression
- NEON SIMD (ARM), scalar fallback (x86)
- Multi-threaded matmul
- Top-p sampling with temperature

What's excluded (use full quant.cpp for these):
- Metal/CUDA/Vulkan GPU backends
- MoE routing
- Polar/QJL/Turbo quantization types

Tested: SmolLM2-1.7B produces "Paris, a city rich in history"
for "The capital of France is" prompt. Qwen3.5-0.8B also works.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: unamedkr

examples/single_header_example.c

@@ -0,0 +1,37 @@
+#define QUANT_IMPLEMENTATION
+#include "../quant.h"
+#include <stdio.h>
+
+static void print_token(const char* text, void* ud) {
+    (void)ud;
+    printf("%s", text);
+    fflush(stdout);
+}
+
+int main(int argc, char** argv) {
+    if (argc < 2) {
+        fprintf(stderr, "Usage: %s <model.gguf> [prompt]\n", argv[0]);
+        return 1;
+    }
+    const char* prompt = argc > 2 ? argv[2] : "Hello!";
+
+    quant_model* m = quant_load(argv[1]);
+    if (!m) { fprintf(stderr, "Failed to load model\n"); return 1; }
+
+    quant_config cfg = {
+        .temperature = 0.7f,
+        .top_p = 0.9f,
+        .max_tokens = 100,
+        .n_threads = 4,
+        .kv_compress = 1  // 4-bit KV compression
+    };
+    quant_ctx* c = quant_new(m, &cfg);
+
+    printf("Prompt: %s\n---\n", prompt);
+    quant_generate(c, prompt, print_token, NULL);
+    printf("\n---\n");
+
+    quant_free_ctx(c);
+    quant_free_model(m);
+    return 0;
+}