Remove old unused CPU graph building path from idx_compute_scores_tile.

Author: createthis

src/llama-sparse-indexer.cpp

@@ -60,53 +60,7 @@ ggml_tensor * sparse_attn_indexer::idx_compute_scores_tile(
         t0_us = ggml_time_us();
     }
 
-    ggml_tensor * scores_acc = nullptr;
-    long HEAD_CHUNK = H;
-
-    if (const char *env = getenv("LLAMA_SPARSE_TOPK_HEAD_CHUNK")) {
-        long v = strtol(env, nullptr, 10);
-        if (v > 0) HEAD_CHUNK = v;
-    }
-    if (HEAD_CHUNK > (long)H) HEAD_CHUNK = (long)H;
-    if (HEAD_CHUNK < 1) HEAD_CHUNK = 1;
-
-    ggml_tensor * w_slice = ggml_view_2d(ctx, weights, H, Tc, weights->nb[1], t0*weights->nb[1]);
-
-    for (int64_t h0 = 0; h0 < H; h0 += HEAD_CHUNK) {
-        int64_t ch = std::min<int64_t>(HEAD_CHUNK, H - h0);
-        size_t q_off_head = (size_t)t0 * q3d->nb[1] + (size_t)h0 * q3d->nb[2];
-        ggml_tensor * q_chunk_3d = ggml_view_3d(ctx, q3d, D, Tc, ch, q3d->nb[1], q3d->nb[2], q_off_head);
-        // permute [D, Tc, ch] -> [D, ch, Tc] so that head index is contiguous inside token-major layout
-        ggml_tensor * q_chunk_ht = ggml_permute(ctx, q_chunk_3d, 0, 2, 1, 3); // [D, ch, Tc]
-        q_chunk_ht = ggml_cont(ctx, q_chunk_ht);
-        ggml_tensor * q_chunk_2d = ggml_reshape_2d(ctx, q_chunk_ht, D, Tc*ch);
-        ggml_tensor * b_q = q_chunk_2d;
-        if (use_fp16 && q_chunk_2d->type != GGML_TYPE_F16) {
-            b_q = ggml_cast(ctx, q_chunk_2d, GGML_TYPE_F16);
-            b_q = ggml_cont(ctx, b_q);
-        }
-        ggml_tensor * k_slice = ggml_view_2d(ctx, a_k, D, kv_end, a_k->nb[1], 0);
-        ggml_tensor * logits_chunk = ggml_mul_mat(ctx, k_slice, b_q); // [kv_end, Tc*ch]
-        logits_chunk = ggml_cont(ctx, logits_chunk);
-        ggml_tensor * logits_chunk_3d = ggml_reshape_3d(ctx, logits_chunk, kv_end, ch, Tc); // [kv_end, ch, Tc]
-        logits_chunk_3d = ggml_relu(ctx, logits_chunk_3d);
-        size_t w_off_chunk = (size_t)h0 * w_slice->nb[0];
-        ggml_tensor * w_sub_2d = ggml_view_2d(ctx, w_slice, ch, Tc, w_slice->nb[1], w_off_chunk); // [ch, Tc]
-        w_sub_2d = ggml_cont(ctx, w_sub_2d);
-        ggml_tensor * w_sub_3d = ggml_reshape_3d(ctx, w_sub_2d, ch, 1, Tc); // [ch,1,Tc]
-        ggml_tensor * log_p = ggml_permute(ctx, logits_chunk_3d, 1, 0, 2, 3); // [ch, N_kv, Tc]
-        log_p = ggml_cont(ctx, log_p);
-        ggml_tensor * w_bc  = ggml_repeat(ctx, w_sub_3d, log_p);             // [ch, N_kv, Tc]
-        w_bc = ggml_cont(ctx, w_bc);
-        ggml_tensor * prod  = ggml_mul(ctx, log_p, w_bc);                     // [ch, N_kv, Tc]
-        ggml_tensor * sum_ch= ggml_sum_rows(ctx, prod);                       // [1, N_kv, Tc]
-        // sum_ch is [1, N_kv, Tc] with linear layout kv-major, then t; reshape directly to [kv_end, Tc]
-        ggml_tensor * scores_chunk = ggml_reshape_2d(ctx, sum_ch, kv_end, Tc);   // [kv_end, Tc]
-        scores_acc = scores_acc ? ggml_add(ctx, scores_acc, scores_chunk) : scores_chunk;
-    }
-    ggml_tensor * scores_tc = scores_acc;
-
-    // CPU reference for TL FP8 lightning indexer, using GGML FP8 encode/decode.
+    // CPU FP8 Lightning Indexer reference, using GGML FP8 helpers.
     // Layout conventions:
     //   q3d : [D, T_total, H]
     //   a_k : [D, N_kv]
@@ -212,7 +166,7 @@ ggml_tensor * sparse_attn_indexer::idx_compute_scores_tile(
     }
 
     // Materialize scores_tc as a new F32 tensor [kv_end, Tc]
-    scores_tc = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, kv, Tc);
+    ggml_tensor * scores_tc = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, kv, Tc);
     std::memcpy(scores_tc->data, out.data(), out.size() * sizeof(float));
     scores_tc->op = GGML_OP_NONE;