Merge branch 'main' into gpu_integration

Author: ColoCarletti

.github/workflows/benchmark-gpu.yml

@@ -33,7 +33,7 @@ concurrency:
   cancel-in-progress: true
 
 env:
-  # Vast offer search: RTX 5090, >=16 cores, >=64GB RAM, >=64GB disk, verified +
+  # Vast offer search: RTX 5090, >=16 cores, >=96GB RAM, >=64GB disk, verified +
   # rentable, Blackwell-capable driver, <= cap.
   GPU_NAME: RTX_5090
   PRICE_CAP: "1"
@@ -169,7 +169,7 @@ jobs:
           MIN_DRIVER: "580"
         run: |
           # cpu_ram filter is in GB.
-          QUERY="gpu_name=${GPU_NAME} num_gpus=1 cpu_cores_effective>=16 cpu_ram>=64 disk_space>=64 verified=true rentable=true cuda_max_good>=12.8 dph_total<=${PRICE_CAP}"
+          QUERY="gpu_name=${GPU_NAME} num_gpus=1 cpu_cores_effective>=16 cpu_ram>=96 disk_space>=64 verified=true rentable=true cuda_max_good>=12.8 dph_total<=${PRICE_CAP}"
           echo "Query: $QUERY (+ client-side driver_version major >= $MIN_DRIVER)"
           # Keep only offers whose driver major >= MIN_DRIVER, then most expensive first
           # (within the price cap) — premium hosts have faster disks/network (quicker image
@@ -190,7 +190,7 @@ jobs:
             sleep "$OFFER_INTERVAL"
           done
           if [ -z "$OFFER_ID" ]; then
-            echo "::error::No RTX 5090 offer matched after $OFFER_ATTEMPTS attempts (>=16 cores, >=64GB RAM, >=64GB disk, driver>=${MIN_DRIVER}, <= \$${PRICE_CAP}/hr)"
+            echo "::error::No RTX 5090 offer matched after $OFFER_ATTEMPTS attempts (>=16 cores, >=96GB RAM, >=64GB disk, driver>=${MIN_DRIVER}, <= \$${PRICE_CAP}/hr)"
             exit 1
           fi
           echo "id=$OFFER_ID" >> "$GITHUB_OUTPUT"

crypto/math-cuda/tests/merkle_root_parity.rs

@@ -55,9 +55,11 @@ fn gpu_merkle_root(columns: &[Vec<u64>], blowup: usize, weights: &[u64]) -> [u8;
         }
     }
 
-    // Per-row leaves (rows_per_leaf = 1): this parity test compares the generic
-    // keccak-leaves + Merkle primitives against a per-row CPU reference.
-    let gpu_leaves = math_cuda::merkle::keccak_leaves_base(&flat, n_lde, num_cols, n_lde, 1)
+    // Row-pair leaves (rows_per_leaf = 2, matching `ROWS_PER_LEAF`): the CPU
+    // reference is `commit_rows_bit_reversed`, which hashes bit-reversed row
+    // pairs into each leaf, so the generic GPU keccak-leaves + Merkle path must
+    // use the same row-pair layout to produce a matching root.
+    let gpu_leaves = math_cuda::merkle::keccak_leaves_base(&flat, n_lde, num_cols, n_lde, 2)
         .expect("GPU keccak leaves");
     let nodes =
         math_cuda::merkle::build_merkle_tree_on_device(&gpu_leaves).expect("GPU Merkle tree");
@@ -191,8 +193,10 @@ fn gpu_ext3_merkle_root(columns: &[Vec<Fp3>], blowup: usize, weights: &[u64]) ->
         }
     }
 
+    // Row-pair leaves (rows_per_leaf = 2, matching `ROWS_PER_LEAF`) to match the
+    // row-pair `commit_rows_bit_reversed` CPU reference below.
     let gpu_leaves =
-        math_cuda::merkle::keccak_leaves_ext3(&flat_for_keccak, lde_size, num_cols, lde_size, 1)
+        math_cuda::merkle::keccak_leaves_ext3(&flat_for_keccak, lde_size, num_cols, lde_size, 2)
             .expect("GPU ext3 keccak leaves");
     let nodes =
         math_cuda::merkle::build_merkle_tree_on_device(&gpu_leaves).expect("GPU Merkle tree");