Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
attn: vectorize the MLA absorption helpers (qt_addrow / qt_matvec_rows int4)
veerareddyvishal144a66c99a7e00c
diff --git a/c/colibri.c b/c/colibri.cindex 4524c105..27620d25 100644--- a/c/colibri.c+++ b/c/colibri.c@@ -2025,6 +2025,9 @@ static void qt_addrow(const QT *t, int row, float coef, float *acc){float c=coef*t->s[row];if(t->fmt==1){ const int8_t *w=t->q8+(int64_t)row*I; for(int i=0;i<I;i++) acc[i]+=c*(float)w[i]; return; }if(t->fmt==2){ const uint8_t *w=t->q4+(int64_t)row*((I+1)/2);+#if defined(__AVX512F__) && defined(__AVX512BW__)+ axpy_i4f_avx512(w,c,acc,I); return; /* bit-identical: one fma per element */+#endiffor(int i=0;i+1<I;i+=2){ uint8_t b=w[i>>1]; acc[i]+=c*((int)(b&0xF)-8); acc[i+1]+=c*((int)(b>>4)-8); }if(I&1){ uint8_t b=w[I>>1]; acc[I-1]+=c*((int)(b&0xF)-8); } return; }const uint8_t *w=t->q4+(int64_t)row*((I+3)/4);@@ -2045,6 +2048,10 @@ static void qt_matvec_rows(const QT *t, int r0, int n, const float *x, float *y)else if(t->fmt==1){ const int8_t *w=t->q8+(int64_t)row*I; float s=t->s[row];float acc=0; for(int i=0;i<I;i++) acc+=(float)w[i]*x[i]; a=acc*s; }else if(t->fmt==2){ const uint8_t *w=t->q4+(int64_t)row*((I+1)/2); float s=t->s[row]; float acc=0;+#if defined(__AVX512F__) && defined(__AVX512BW__)+ /* same accumulation-order tradeoff (and gate) as matmul_i4's I4_ACC512 path */+ if(g_i4_acc512 && !(I&31)){ y[j]=dot_i4f_avx512(w,x,I)*s; continue; }+#endiffor(int i=0;i+1<I;i+=2){ uint8_t b=w[i>>1]; acc+=((int)(b&0xF)-8)*x[i]+((int)(b>>4)-8)*x[i+1]; }if(I&1){ uint8_t b=w[I>>1]; acc+=((int)(b&0xF)-8)*x[I-1]; } a=acc*s; }else if(t->fmt==4){ /* per-gruppo, come matmul_i4_grouped / per-group, as matmul_i4_grouped */diff --git a/c/quant.h b/c/quant.hindex b3755380..d00a39dc 100644--- a/c/quant.h+++ b/c/quant.h@@ -59,6 +59,24 @@ static inline float dot_i4f_avx512(const uint8_t *w,const float *x,int I){}return _mm512_reduce_add_ps(_mm512_add_ps(acc0,acc1));}+/* acc[0..I) += coef * dequant(int4 row) — the axpy twin of dot_i4f_avx512, for the+ * MLA absorption path (qt_addrow). Each acc[i] receives exactly ONE fma per call+ * (no cross-element accumulation), so this is bit-identical to the scalar loop+ * (gcc -O3 -ffp-contract already emits scalar fma there). Tail handled scalar. */+static inline void axpy_i4f_avx512(const uint8_t *w,float coef,float *acc,int I){+ const __m128i m4=_mm_set1_epi8(0x0F); const __m512i b8=_mm512_set1_epi32(8);+ const __m512 cv=_mm512_set1_ps(coef); int i=0;+ for(;i+32<=I;i+=32){ __m128i by=_mm_loadu_si128((const __m128i*)(w+(i>>1)));+ __m128i lo=_mm_and_si128(by,m4),hi=_mm_and_si128(_mm_srli_epi16(by,4),m4);+ __m128i n0=_mm_unpacklo_epi8(lo,hi),n1=_mm_unpackhi_epi8(lo,hi);+ __m512 w0=_mm512_cvtepi32_ps(_mm512_sub_epi32(_mm512_cvtepu8_epi32(n0),b8));+ __m512 w1=_mm512_cvtepi32_ps(_mm512_sub_epi32(_mm512_cvtepu8_epi32(n1),b8));+ _mm512_storeu_ps(acc+i, _mm512_fmadd_ps(cv,w0,_mm512_loadu_ps(acc+i)));+ _mm512_storeu_ps(acc+i+16,_mm512_fmadd_ps(cv,w1,_mm512_loadu_ps(acc+i+16)));+ }+ for(;i+1<I;i+=2){ uint8_t b=w[i>>1]; acc[i]+=coef*(float)((int)(b&0xF)-8); acc[i+1]+=coef*(float)((int)(b>>4)-8); }+ if(i<I){ uint8_t b=w[i>>1]; acc[i]+=coef*(float)((int)(b&0xF)-8); }+}static int i4_acc512_selftest(void){enum { N=224 }; uint8_t w[(N+1)/2]; float x[N];for(int i=0;i<N;i++){