cuda: GDN decode conv-chain fusion and alpha/beta mat-vec pair on by default (GGML_CUDA_GDN_FUSION=0 disables), bit-identical, +7.9% 9B decode on H200

ggml/src/ggml-cuda/common.cuh

@@ -1597,10 +1597,11 @@ struct ggml_cuda_gdn_gather_context {
     }
 };
 
-// GDN decode conv chain fused into one launch (GGML_CUDA_GDN_FUSION=1): the conv-state GET_ROWS is skipped and
-// registered under the CONCAT it feeds; the CONCAT launches the fused kernel, which also computes the conv-state CPY,
-// SSM_CONV + SILU and the q/k L2_NORM, and those nodes are marked done. Keyed by node identity, valid for one graph
-// evaluation, owned by the evaluating context (reset with the other per-evaluation registries).
+// GDN decode conv chain fused into one launch (on by default, GGML_CUDA_GDN_FUSION=0 disables it): the conv-state
+// GET_ROWS is skipped and registered under the CONCAT it feeds; the CONCAT launches the fused kernel, which also
+// computes the conv-state CPY, SSM_CONV + SILU and the q/k L2_NORM, and those nodes are marked done. Keyed by node
+// identity, valid for one graph evaluation, owned by the evaluating context (reset with the other per-evaluation
+// registries).
 struct ggml_cuda_gdn_conv_context {
     std::unordered_map<const ggml_tensor *, const ggml_tensor *> gathers; // CONCAT -> the GET_ROWS it replaces
     std::unordered_set<const ggml_tensor *>                      done;    // nodes computed by a fused launch

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -3066,7 +3066,7 @@ static int ggml_cuda_try_fwht_q8(ggml_backend_cuda_context & ctx, const ggml_cgr
     return consumed;
 }
 
-// GDN conv chain (GGML_CUDA_GDN_FUSION=1, off by default). For n_t tokens of one sequence (a decode step, n_t = 1, or a
+// GDN conv chain (on by default, GGML_CUDA_GDN_FUSION=0 disables it). For n_t tokens of one sequence (a decode step, n_t = 1, or a
 // speculative verify step, n_t = 2..GGML_CUDA_GDN_CONV_MAX_TOKENS), a hybrid-model gated delta net layer runs these
 // launch-bound kernels between the qkv projection and the GDN op:
 //   GET_ROWS(conv cache, s_copy) -> RESHAPE -> CONCAT(state, qkv^T)
@@ -3080,8 +3080,10 @@ static int ggml_cuda_try_fwht_q8(ggml_backend_cuda_context & ctx, const ggml_cgr
 // of it) must not touch any memory the fused kernel reads or writes. Megakernel-style launch removal in the spirit of
 // Hazy Research's "no bubbles" low-latency decode work, applied as a graph-level fusion rather than a persistent kernel.
 static bool ggml_cuda_gdn_fusion_enabled() {
-    static const bool enabled = getenv("GGML_CUDA_GDN_FUSION") != nullptr &&
-                                atoi(getenv("GGML_CUDA_GDN_FUSION")) != 0;
+    static const bool enabled = [] {
+        const char * env = getenv("GGML_CUDA_GDN_FUSION");
+        return env == nullptr || atoi(env) != 0;
+    }();
     return enabled;
 }
 
@@ -4218,7 +4220,8 @@ static int ggml_cuda_try_batched_cpy(ggml_backend_cuda_context * cuda_ctx, ggml_
 }
 
 // Two consecutive independent F32/F16/BF16 mat-vecs on the same activation (the GDN alpha and beta projections, which
-// the fused GDN op consumes raw) as one launch, GGML_CUDA_GDN_FUSION=1. Returns the number of nodes consumed.
+// the fused GDN op consumes raw) as one launch (on by default, GGML_CUDA_GDN_FUSION=0 disables it). Returns the
+// number of nodes consumed.
 static int ggml_cuda_try_mmvf_pair(ggml_backend_cuda_context & ctx, ggml_cgraph * cgraph, int i) {
     ggml_tensor * mm0 = cgraph->nodes[i];
     int j = i + 1;

tests/test-backend-ops.cpp

@@ -4140,7 +4140,7 @@ struct test_ssm_conv_bias_silu : public test_case {
 // GET_ROWS (conv state row) -> CONCAT with the new tokens -> CPY of the last window back into a cache row (n_rs == 0),
 // or one CPY per rollback snapshot slot (n_rs > 0: 1 + n_rs slots, as llm_build_delta_net_base::build_conv_state writes
 // them), SSM_CONV -> SILU -> L2_NORM of the leading q/k heads of every token. The CUDA backend can run it as one
-// launch (GGML_CUDA_GDN_FUSION=1); the L2_NORM, the SILU output and the written cache rows are all checked.
+// launch (by default; GGML_CUDA_GDN_FUSION=0 disables it); the L2_NORM, the SILU output and the written cache rows are all checked.
 struct test_gdn_conv_chain : public test_case {
     const int64_t d_conv;
     const int64_t head_dim;  // L2 row length
@@ -4230,7 +4230,7 @@ struct test_gdn_conv_chain : public test_case {
 };
 
 // Two mat-vecs of one activation with same-shape float weights (the GDN alpha and beta projections); the CUDA backend
-// can run them as one launch (GGML_CUDA_GDN_FUSION=1). Both products are checked.
+// can run them as one launch (by default; GGML_CUDA_GDN_FUSION=0 disables it). Both products are checked.
 struct test_mul_mat_vec_pair : public test_case {
     const ggml_type type;
     const int64_t   k;