v6_simple.c File Reference

Simplified v6 CLI using only generic kernels. More...

#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include <string.h>
#include <time.h>
#include <math.h>
#include <omp.h>
#include "ckernel_engine.h"

Go to the source code of this file.

Macros
#define	_GNU_SOURCE
#define	ALIGN_CONTEXT   32768
#define	ALIGN_EMBED   896
#define	ALIGN_HEAD   64
#define	MODEL_EMBED_DIM   896
#define	MODEL_HEAD_DIM   64
#define	MODEL_INTERMEDIATE   4864
#define	MODEL_INTERMEDIATE_SIZE   4864
#define	MODEL_MAX_SEQ_LEN   32768
#define	MODEL_NUM_HEADS   14
#define	MODEL_NUM_KV_HEADS   2
#define	MODEL_NUM_LAYERS   24
#define	MODEL_VOCAB_SIZE   128256

Functions
static void	apply_rope (float *x, int seq_len, int head_dim)
static void	gemm_nt (const float *input, const float *weight, float *output, int rows, int cols, int common)
int	main (int argc, char **argv)
static void	residual_add (float *residual, float *addend, int n)
static void	silu (float *x, int n)
static void	simple_attention (const float *q, const float *k, const float *v, float *output, int num_heads, int num_kv_heads, int seq_len, int head_dim)
static void	simple_embedding (const int32_t *tokens, int num_tokens, const float *weight, float *output, int vocab_size, int embed_dim)
static void	simple_rmsnorm (const float *input, const float *gamma, float *output, int tokens, int d_model, float eps)
static void	softmax (float *x, int n)
void	v6_prefill (const float *embed_weight, const int32_t *tokens, int num_tokens, float *logits)

Detailed Description

Simplified v6 CLI using only generic kernels.

This is a minimal v6 implementation that uses:

• Generic GEMM (gemm_blocked_serial) instead of quantized kernels
• Generic RMSNorm
• Precomputed RoPE
• OMP parallelization for prefill

Definition in file v6_simple.c.

Macro Definition Documentation

_GNU_SOURCE

#define _GNU_SOURCE

Definition at line 12 of file v6_simple.c.

ALIGN_CONTEXT

#define ALIGN_CONTEXT   32768

Definition at line 37 of file v6_simple.c.

ALIGN_EMBED

#define ALIGN_EMBED   896

Definition at line 34 of file v6_simple.c.

ALIGN_HEAD

#define ALIGN_HEAD   64

Definition at line 35 of file v6_simple.c.

MODEL_EMBED_DIM

#define MODEL_EMBED_DIM   896

Definition at line 24 of file v6_simple.c.

MODEL_HEAD_DIM

#define MODEL_HEAD_DIM   64

Definition at line 28 of file v6_simple.c.

MODEL_INTERMEDIATE

#define MODEL_INTERMEDIATE   4864

Definition at line 36 of file v6_simple.c.

MODEL_INTERMEDIATE_SIZE

#define MODEL_INTERMEDIATE_SIZE   4864

Definition at line 29 of file v6_simple.c.

MODEL_MAX_SEQ_LEN

#define MODEL_MAX_SEQ_LEN   32768

Definition at line 31 of file v6_simple.c.

MODEL_NUM_HEADS

#define MODEL_NUM_HEADS   14

Definition at line 26 of file v6_simple.c.

MODEL_NUM_KV_HEADS

#define MODEL_NUM_KV_HEADS   2

Definition at line 27 of file v6_simple.c.

MODEL_NUM_LAYERS

#define MODEL_NUM_LAYERS   24

Definition at line 25 of file v6_simple.c.

MODEL_VOCAB_SIZE

#define MODEL_VOCAB_SIZE   128256

Definition at line 30 of file v6_simple.c.

Function Documentation

apply_rope()

static void apply_rope ( float *  x, int  seq_len, int  head_dim  )

static

Definition at line 173 of file v6_simple.c.

                                                            {
    /* Simplified - just identity for now */
    (void)x;
    (void)seq_len;
    (void)head_dim;
}

Referenced by v6_prefill().

gemm_nt()

static void gemm_nt ( const float *  input, const float *  weight, float *  output, int  rows, int  cols, int  common  )

static

Definition at line 145 of file v6_simple.c.

                                                    {
    for (int r = 0; r < rows; r++) {
        for (int c = 0; c < cols; c++) {
            float sum = 0.0f;
            for (int k = 0; k < common; k++) {
                sum += input[r * common + k] * weight[c * common + k];
            }
            output[r * cols + c] = sum;
        }
    }
}

Referenced by v6_prefill().

main()

int main	(	int	argc,
		char **	argv
	)

Definition at line 295 of file v6_simple.c.

                                {
    printf("=== V6 Simple CLI ===\n");
    printf("Generic kernel implementation\n");
    printf("OMP parallelization for prefill\n\n");
 
    if (argc < 2) {
        printf("Usage: %s <weights.bin> [options]\n", argv[0]);
        printf("\nOptions:\n");
        printf("  -p, --prompt <text>   Input prompt\n");
        printf("  -t, --tokens <n>      Max tokens (default: 50)\n");
        printf("  -h, --help            Show help\n");
        return 1;
    }
 
    const char *weights_path = argv[1];
    const char *prompt = "Hello";
    int max_tokens = 50;
 
    for (int i = 2; i < argc; i++) {
        if (strcmp(argv[i], "-p") == 0 || strcmp(argv[i], "--prompt") == 0) {
            prompt = argv[++i];
        } else if (strcmp(argv[i], "-t") == 0 || strcmp(argv[i], "--tokens") == 0) {
            max_tokens = atoi(argv[++i]);
        } else if (strcmp(argv[i], "-h") == 0 || strcmp(argv[i], "--help") == 0) {
            printf("Usage: %s <weights.bin> [options]\n", argv[0]);
            return 0;
        }
    }
 
    printf("Model: Qwen2 0.5B (generic kernels)\n");
    printf("Prompt: %s\n", prompt);
    printf("Max tokens: %d\n", max_tokens);
    printf("\n[Note: This is a simplified v6 implementation using generic kernels]\n");
    printf("[Real weights loading and inference would require full implementation]\n");
 
    /* Placeholder for actual inference */
    printf("\nAssistant: (v6 placeholder - full implementation pending)\n");
 
    return 0;
}

residual_add()

static void residual_add ( float *  residual, float *  addend, int  n  )

static

Definition at line 166 of file v6_simple.c.

                                                                {
    for (int i = 0; i < n; i++) {
        residual[i] += addend[i];
    }
}

Referenced by v6_prefill().

silu()

static void silu ( float *  x, int  n  )

static

Definition at line 159 of file v6_simple.c.

                                  {
    for (int i = 0; i < n; i++) {
        x[i] = x[i] / (1.0f + expf(-x[i]));
    }
}

Referenced by v6_prefill().

simple_attention()

static void simple_attention ( const float *  q, const float *  k, const float *  v, float *  output, int  num_heads, int  num_kv_heads, int  seq_len, int  head_dim  )

static

Definition at line 80 of file v6_simple.c.

                                                        {
    int hidden_dim = num_heads * head_dim;
 
    /* For each head, compute attention */
    for (int h = 0; h < num_heads; h++) {
        const float *q_head = q + h * head_dim;
        const float *k_head = k;  /* KV heads repeated */
        const float *v_head = v;
 
        /* Repeat K/V for GQA */
        int repeat = num_heads / num_kv_heads;
 
        float *out_head = output + h * head_dim;
 
        /* Compute attention scores */
        float scores[MODEL_MAX_SEQ_LEN];
        for (int t = 0; t < seq_len; t++) {
            float score = 0.0f;
            for (int d = 0; d < head_dim; d++) {
                score += q_head[d] * k_head[t * head_dim + d];
            }
            scores[t] = score / sqrtf((float)head_dim);
        }
 
        /* Causal mask */
        for (int t = 0; t < seq_len; t++) {
            if (t >= seq_len - 1) {
                scores[t] = scores[t];  /* Last token can attend to all */
            } else {
                scores[t] = -1e9f;  /* Mask future tokens */
            }
        }
 
        /* Softmax */
        softmax(scores, seq_len);
 
        /* Weighted sum */
        for (int d = 0; d < head_dim; d++) {
            float sum = 0.0f;
            for (int t = 0; t < seq_len; t++) {
                sum += scores[t] * v_head[t * head_dim + d];
            }
            out_head[d] = sum;
        }
    }
}

References MODEL_MAX_SEQ_LEN, score, and softmax().

Referenced by v6_prefill().

simple_embedding()

static void simple_embedding ( const int32_t *  tokens, int  num_tokens, const float *  weight, float *  output, int  vocab_size, int  embed_dim  )

static

Definition at line 130 of file v6_simple.c.

                                                           {
    for (int t = 0; t < num_tokens; t++) {
        int token_id = tokens[t];
        if (token_id >= 0 && token_id < vocab_size) {
            memcpy(output + t * embed_dim, weight + token_id * embed_dim,
                   embed_dim * sizeof(float));
        } else {
            memset(output + t * embed_dim, 0, embed_dim * sizeof(float));
        }
    }
}

References vocab_size.

Referenced by v6_prefill().

simple_rmsnorm()

static void simple_rmsnorm ( const float *  input, const float *  gamma, float *  output, int  tokens, int  d_model, float  eps  )

static

Definition at line 40 of file v6_simple.c.

                                                               {
    for (int t = 0; t < tokens; t++) {
        const float *in_row = input + t * d_model;
        float *out_row = output + t * d_model;
 
        /* Compute variance */
        float variance = 0.0f;
        for (int i = 0; i < d_model; i++) {
            variance += in_row[i] * in_row[i];
        }
        variance /= d_model;
 
        /* Normalize */
        float scale = 1.0f / sqrtf(variance + eps);
        for (int i = 0; i < d_model; i++) {
            out_row[i] = in_row[i] * gamma[i] * scale;
        }
    }
}

Referenced by v6_prefill().

softmax()

static void softmax ( float *  x, int  n  )

static

Definition at line 62 of file v6_simple.c.

                                     {
    float max_val = x[0];
    for (int i = 1; i < n; i++) {
        if (x[i] > max_val) max_val = x[i];
    }
 
    float sum = 0.0f;
    for (int i = 0; i < n; i++) {
        x[i] = expf(x[i] - max_val);
        sum += x[i];
    }
 
    for (int i = 0; i < n; i++) {
        x[i] /= sum;
    }
}

Referenced by simple_attention().

v6_prefill()

void v6_prefill	(	const float *	embed_weight,
		const int32_t *	tokens,
		int	num_tokens,
		float *	logits
	)

Definition at line 181 of file v6_simple.c.

                               {
    if (!embed_weight || !tokens || num_tokens <= 0) return;
 
    /* Allocate buffers */
    const int embed_dim = ALIGN_EMBED;
    const int intermediate = MODEL_INTERMEDIATE;
    const int num_layers = MODEL_NUM_LAYERS;
    const int num_heads = MODEL_NUM_HEADS;
    const int num_kv_heads = MODEL_NUM_KV_HEADS;
    const int head_dim = MODEL_HEAD_DIM;
 
    /* Per-token hidden states: (num_tokens) x (num_layers + 1) x embed_dim */
    float *hidden = malloc(num_tokens * (num_layers + 1) * embed_dim * sizeof(float));
    if (!hidden) {
        fprintf(stderr, "Failed to allocate hidden states\n");
        return;
    }
 
    /* Temporary buffers per layer */
    float *q = malloc(num_heads * head_dim * sizeof(float));
    float *k = malloc(num_kv_heads * head_dim * sizeof(float));
    float *v = malloc(num_kv_heads * head_dim * sizeof(float));
    float *attn = malloc(num_heads * head_dim * sizeof(float));
    float *mlp = malloc(intermediate * sizeof(float));
 
    if (!q || !k || !v || !attn || !mlp) {
        fprintf(stderr, "Failed to allocate temp buffers\n");
        free(hidden);
        free(q);
        free(k);
        free(v);
        free(attn);
        free(mlp);
        return;
    }
 
    /* Dummy layer weights (in real impl, these come from mapped memory) */
    const float *ln1_gamma = NULL;  /* Would come from weights */
    const float *ln2_gamma = NULL;
    const float *wq = NULL, *wk = NULL, *wv = NULL, *wo = NULL;
    const float *w1 = NULL, *w2 = NULL;
 
    /* OMP parallel for over tokens */
    #pragma omp parallel for schedule(dynamic, 1)
    for (int t = 0; t < num_tokens; t++) {
        float *h = hidden + t * (num_layers + 1) * embed_dim;
 
        /* Embedding lookup */
        simple_embedding(tokens + t, 1, embed_weight, h, MODEL_VOCAB_SIZE, embed_dim);
 
        /* Process through layers */
        for (int layer = 0; layer < num_layers; layer++) {
            float *layer_in = h;
            float *layer_out = h + embed_dim;
 
            /* RMSNorm */
            simple_rmsnorm(layer_in, ln1_gamma, layer_in, 1, embed_dim, 1e-6f);
 
            /* QKV projection */
            gemm_nt(layer_in, wq, q, 1, num_heads * head_dim, embed_dim);
            gemm_nt(layer_in, wk, k, 1, num_kv_heads * head_dim, embed_dim);
            gemm_nt(layer_in, wv, v, 1, num_kv_heads * head_dim, embed_dim);
 
            /* RoPE */
            apply_rope(q, 1, head_dim);
            apply_rope(k, 1, head_dim);
 
            /* Attention */
            simple_attention(q, k, v, attn, num_heads, num_kv_heads, 1, head_dim);
 
            /* Output projection */
            gemm_nt(attn, wo, layer_out, 1, embed_dim, num_heads * head_dim);
 
            /* Residual */
            residual_add(layer_in, layer_out, embed_dim);
 
            /* RMSNorm before MLP */
            simple_rmsnorm(layer_in, ln2_gamma, layer_in, 1, embed_dim, 1e-6f);
 
            /* MLP */
            gemm_nt(layer_in, w1, mlp, 1, 2 * intermediate, embed_dim);
            silu(mlp, 2 * intermediate);
            gemm_nt(mlp, w2, layer_out, 1, embed_dim, intermediate);
 
            /* Residual */
            residual_add(layer_in, layer_out, embed_dim);
        }
 
        /* Copy to output area */
        memcpy(hidden + t * (num_layers + 1) * embed_dim +
               num_layers * embed_dim, h, embed_dim * sizeof(float));
    }
 
    /* Final RMSNorm over all tokens */
    float *final_out = malloc(num_tokens * embed_dim * sizeof(float));
    if (final_out) {
        simple_rmsnorm(hidden + num_layers * embed_dim, ln1_gamma, final_out,
                       num_tokens, embed_dim, 1e-6f);
 
        /* LM head */
        gemm_nt(final_out, embed_weight, logits, num_tokens, MODEL_VOCAB_SIZE, embed_dim);
 
        free(final_out);
    }
 
    free(hidden);
    free(q);
    free(k);
    free(v);
    free(attn);
    free(mlp);
}

References ALIGN_EMBED, apply_rope(), gemm_nt(), MODEL_HEAD_DIM, MODEL_INTERMEDIATE, MODEL_NUM_HEADS, MODEL_NUM_KV_HEADS, MODEL_NUM_LAYERS, MODEL_VOCAB_SIZE, residual_add(), silu(), simple_attention(), simple_embedding(), and simple_rmsnorm().