diff --git a/gpttype_adapter.cpp b/gpttype_adapter.cpp
index b876f8441..5996911a1 100644
--- a/gpttype_adapter.cpp
+++ b/gpttype_adapter.cpp
@@ -563,7 +563,7 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
             rwkv_ctx_v3->logits_out = (float *)malloc(logitbufsiz);
             rwkv_ctx_v3->state_in = nullptr;
 
-            bool testeval = rwkv_eval(rwkv_ctx_v3, 0, rwkv_ctx_v3->state_in, rwkv_ctx_v3->state_out, rwkv_ctx_v3->logits_out);
+            bool testeval = rwkv_eval(rwkv_ctx_v3, params.n_threads, 0, rwkv_ctx_v3->state_in, rwkv_ctx_v3->state_out, rwkv_ctx_v3->logits_out);
             if (!testeval)
             {
                 printf("\nError: RWKV Init Eval Failed!\n");
@@ -1162,12 +1162,12 @@ generation_outputs gpttype_generate(const generation_inputs inputs, generation_o
                 {
                     if(embd.size()>1)
                     {
-                        evalres = rwkv_eval_sequence(rwkv_ctx_v3, (uint32_t*)embd.data(), embd.size(), rwkv_ctx_v3->state_in, rwkv_ctx_v3->state_out, rwkv_ctx_v3->logits_out);
+                        evalres = rwkv_eval_sequence(rwkv_ctx_v3, params.n_threads, (uint32_t*)embd.data(), embd.size(), rwkv_ctx_v3->state_in, rwkv_ctx_v3->state_out, rwkv_ctx_v3->logits_out);
                     }
                     else
                     {
                     bool ignoreLogits = (!startedsampling && ((int)embd_inp.size() > input_consumed + 2));
-                    evalres = rwkv_eval(rwkv_ctx_v3, embd[0], rwkv_ctx_v3->state_in, rwkv_ctx_v3->state_out, ignoreLogits?nullptr:rwkv_ctx_v3->logits_out);
+                    evalres = rwkv_eval(rwkv_ctx_v3, params.n_threads, embd[0], rwkv_ctx_v3->state_in, rwkv_ctx_v3->state_out, ignoreLogits?nullptr:rwkv_ctx_v3->logits_out);
                     }
 
                     memcpy(logits.data(), rwkv_ctx_v3->logits_out, sizeof(float) * rwkv_vocab.size());
diff --git a/otherarch/gpt2_v3.cpp b/otherarch/gpt2_v3.cpp
index b507357c4..2e7806d3d 100644
--- a/otherarch/gpt2_v3.cpp
+++ b/otherarch/gpt2_v3.cpp
@@ -447,7 +447,6 @@ bool gpt2_eval(
 
     struct ggml_context * ctx0 = ggml_init(params);
     struct ggml_cgraph gf = {};
-    gf.n_threads = n_threads;
 
     struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
     memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
@@ -708,7 +707,7 @@ bool gpt2_eval(
 
     // run the computation
     ggml_build_forward_expand(&gf, inpL);
-    ggml_graph_compute       (ctx0, &gf);
+    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
 
     //if (n_past%100 == 0) {
     //    ggml_graph_print   (&gf);
diff --git a/otherarch/gptj_v3.cpp b/otherarch/gptj_v3.cpp
index be4ad60da..aeaa7bbd5 100644
--- a/otherarch/gptj_v3.cpp
+++ b/otherarch/gptj_v3.cpp
@@ -445,7 +445,6 @@ bool gptj_eval(
 
     struct ggml_context * ctx0 = ggml_init(params);
     struct ggml_cgraph gf = {};
-    gf.n_threads = n_threads;
 
     struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
     memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
@@ -620,7 +619,7 @@ bool gptj_eval(
 
     // run the computation
     ggml_build_forward_expand(&gf, inpL);
-    ggml_graph_compute       (ctx0, &gf);
+    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
 
     //if (n_past%100 == 0) {
     //    ggml_graph_print   (&gf);
diff --git a/otherarch/mpt_v3.cpp b/otherarch/mpt_v3.cpp
index e4cf99fd7..35006688d 100644
--- a/otherarch/mpt_v3.cpp
+++ b/otherarch/mpt_v3.cpp
@@ -383,7 +383,6 @@ bool mpt_eval(const mpt_model & model, const int n_threads, const int n_past,
 
     struct ggml_context * ctx0 = ggml_init(params);
     struct ggml_cgraph gf = {};
-    gf.n_threads = n_threads;
 
     struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
     memcpy(embd->data, embd_inp.data(), N * ggml_element_size(embd));
@@ -543,7 +542,7 @@ bool mpt_eval(const mpt_model & model, const int n_threads, const int n_past,
 
     // run the computation
     ggml_build_forward_expand(&gf, inpL);
-    ggml_graph_compute(ctx0, &gf);
+    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
 
     // std::cout << "Qcur" << std::endl;
     // print_tensor(Qcur);
diff --git a/otherarch/neox_v3.cpp b/otherarch/neox_v3.cpp
index 29b2d06c8..9c1ab2545 100644
--- a/otherarch/neox_v3.cpp
+++ b/otherarch/neox_v3.cpp
@@ -461,7 +461,6 @@ bool gpt_neox_eval(
 
     struct ggml_context * ctx0 = ggml_init(params);
     struct ggml_cgraph gf = {};
-    gf.n_threads = n_threads;
 
     struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
     memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
@@ -639,7 +638,7 @@ bool gpt_neox_eval(
 
     // run the computation
     ggml_build_forward_expand(&gf, inpL);
-    ggml_graph_compute       (ctx0, &gf);
+    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
 
     //if (n_past%100 == 0) {
     //    ggml_graph_print   (&gf);
diff --git a/otherarch/rwkv_v3.cpp b/otherarch/rwkv_v3.cpp
index 0ccf8d8fa..2ef44dd1b 100644
--- a/otherarch/rwkv_v3.cpp
+++ b/otherarch/rwkv_v3.cpp
@@ -1511,7 +1511,6 @@ struct rwkv_context * rwkv_new_context_impl(std::shared_ptr<struct rwkv_instance
     serial_graph.tokens = ggml_new_i32(serial_graph.ctx.ctx, 0);
     serial_graph.cgraph.reset(new(std::nothrow) struct ggml_cgraph());
     RWKV_ASSERT_NULL_MSG(RWKV_ERROR_ALLOC, serial_graph.cgraph, "Failed to allocate serial graph");
-    serial_graph.cgraph->n_threads = n_threads;
 
     RWKV_ASSERT_NULL(RWKV_ERROR_GRAPH, rwkv_build_serial_graph(
         serial_graph.ctx.ctx, instance->model,
@@ -1609,7 +1608,7 @@ void rwkv_get_outputs(const struct rwkv_context * ctx, float * state_out, float
     }
 }
 
-bool rwkv_eval(struct rwkv_context * ctx, const uint32_t token, const float * state_in, float * state_out, float * logits_out) {
+bool rwkv_eval(struct rwkv_context * ctx, const int n_threads, const uint32_t token, const float * state_in, float * state_out, float * logits_out) {
     ctx->last_error = RWKV_ERROR_NONE;
 
     const struct rwkv_file_header & header = ctx->instance->model.header;
@@ -1628,13 +1627,13 @@ bool rwkv_eval(struct rwkv_context * ctx, const uint32_t token, const float * st
         ctx->serial_graph.cgraph->n_leafs = ctx->serial_graph.post_logits_leafs;
     }
 
-    ggml_graph_compute(ctx->serial_graph.ctx.ctx, ctx->serial_graph.cgraph.get());
+    ggml_graph_compute_with_ctx(ctx->serial_graph.ctx.ctx, ctx->serial_graph.cgraph.get(),n_threads);
     rwkv_get_outputs(ctx, state_out, logits_out);
 
     return true;
 }
 
-bool rwkv_eval_sequence(struct rwkv_context * ctx, const uint32_t * sequence, const size_t sequence_len, const float * state_in, float * state_out, float * logits_out) {
+bool rwkv_eval_sequence(struct rwkv_context * ctx, const int n_threads, const uint32_t * sequence, const size_t sequence_len, const float * state_in, float * state_out, float * logits_out) {
     ctx->last_error = RWKV_ERROR_NONE;
 
     const struct rwkv_file_header & header = ctx->instance->model.header;
@@ -1690,7 +1689,6 @@ bool rwkv_eval_sequence(struct rwkv_context * ctx, const uint32_t * sequence, co
         sequence_graph.tokens = ggml_new_tensor_1d(sequence_graph.ctx.ctx, GGML_TYPE_I32, sequence_len);
         sequence_graph.cgraph.reset(new(std::nothrow) struct ggml_cgraph());
         RWKV_ASSERT_FALSE_MSG(RWKV_ERROR_ALLOC, sequence_graph.cgraph, "Failed to allocate sequence graph");
-        sequence_graph.cgraph->n_threads = 1;
 
         RWKV_ASSERT_FALSE(RWKV_ERROR_GRAPH, rwkv_build_sequence_graph(
             sequence_graph.ctx.ctx, ctx->instance->model,
@@ -1717,7 +1715,7 @@ bool rwkv_eval_sequence(struct rwkv_context * ctx, const uint32_t * sequence, co
             ctx->sequence_graph.cgraph->n_leafs = ctx->sequence_graph.post_logits_leafs;
         }
 
-        ggml_graph_compute(ctx->sequence_graph.ctx.ctx, ctx->sequence_graph.cgraph.get());
+        ggml_graph_compute_with_ctx(ctx->sequence_graph.ctx.ctx, ctx->sequence_graph.cgraph.get(),n_threads);
         rwkv_get_outputs(ctx, state_out, logits_out);
     }
 
diff --git a/otherarch/rwkv_v3.h b/otherarch/rwkv_v3.h
index b24812fc2..b9e0d57e2 100644
--- a/otherarch/rwkv_v3.h
+++ b/otherarch/rwkv_v3.h
@@ -111,7 +111,7 @@ extern "C" {
     // - state_in: FP32 buffer of size rwkv_get_state_len(); or NULL, if this is a first pass.
     // - state_out: FP32 buffer of size rwkv_get_state_len(). This buffer will be written to if non-NULL.
     // - logits_out: FP32 buffer of size rwkv_get_logits_len(). This buffer will be written to if non-NULL.
-    RWKV_API bool rwkv_eval(struct rwkv_context * ctx, const uint32_t token, const float * state_in, float * state_out, float * logits_out);
+    RWKV_API bool rwkv_eval(struct rwkv_context *, const int n_threads, const uint32_t token, const float * state_in, float * state_out, float * logits_out);
 
     // Evaluates the model for a sequence of tokens.
     // Uses a faster algorithm than rwkv_eval if you do not need the state and logits for every token. Best used with batch sizes of 64 or so.
@@ -125,7 +125,7 @@ extern "C" {
     // - state_in: FP32 buffer of size rwkv_get_state_len(), or NULL if this is a first pass.
     // - state_out: FP32 buffer of size rwkv_get_state_len(). This buffer will be written to if non-NULL.
     // - logits_out: FP32 buffer of size rwkv_get_logits_len(). This buffer will be written to if non-NULL.
-    RWKV_API bool rwkv_eval_sequence(struct rwkv_context * ctx, const uint32_t * tokens, size_t sequence_len, const float * state_in, float * state_out, float * logits_out);
+    RWKV_API bool rwkv_eval_sequence(struct rwkv_context * ctx, const int n_threads, const uint32_t * tokens, size_t sequence_len, const float * state_in, float * state_out, float * logits_out);
 
     // Returns the number of tokens in the given model's vocabulary.
     // Useful for telling 20B_tokenizer models (n_vocab = 50277) apart from World models (n_vocab = 65536).