Fix Qwen35 mtp warmup (#1987)

* Use hidden state from prev token from qwen mtp

* Fix Qwen35 MTP warmup

* Cleanup + remove unnecessary crippling performance by not using accept to sample draft token

* Provide API to gtet the model arch string

---------

Co-authored-by: SamuelOliveirads <samueloliveira32df@gmail.com>

common/speculative.cpp

@@ -2347,6 +2347,8 @@ void common_speculative_checkpoint_restore(
     common_speculative_checkpoint_discard(ckpt, ctx);
 }
 
+static bool mtp_model_uses_recurrent_conditioning(const common_speculative_state_mtp & state);
+
 void common_speculative_commit(
         common_speculative * spec,
         llama_context * ctx,
@@ -2364,6 +2366,7 @@ void common_speculative_commit(
     const common_speculative_type spec_type_used = spec->curr_impl != nullptr
         ? spec->curr_impl->type
         : COMMON_SPECULATIVE_TYPE_NONE;
+
     const bool any_rejected = (int) ids.size() - 1 < n_draft;
     std::vector<float> mtp_hidden_state_pre;
 
@@ -2552,6 +2555,20 @@ static void mtp_store_target_hidden(
     stored.assign(hidden, hidden + width);
 }
 
+static bool mtp_model_uses_recurrent_conditioning(const common_speculative_state_mtp & state) {
+    if (state.ctx_mtp == nullptr) {
+        return false;
+    }
+
+    const llama_model * model = llama_get_model(state.ctx_mtp);
+    if (!llama_model_has_recurrent(model)) {
+        return false;
+    }
+
+    std::string arch{llama_model_arch_string(model)};
+    return arch == "qwen35" || arch == "qwen35moe";
+}
+
 static void mtp_clear_target_hidden(common_speculative_state_mtp & state, llama_seq_id seq_id) {
     state.target_hidden_by_seq.erase(seq_id);
     state.draft_cache_by_seq.erase(seq_id);
@@ -2726,6 +2743,12 @@ int32_t common_speculative_on_target_batch(
         return 0;
     }
 
+    if (features.width != mtp_state->n_embd) {
+        LOG_ERR("%s: MTP feature width mismatch: got %d expected %d\n",
+                __func__, features.width, mtp_state->n_embd);
+        return -1;
+    }
+
     if (batch.n_seq_id == nullptr || batch.seq_id == nullptr || batch.n_seq_id[0] <= 0 || batch.seq_id[0] == nullptr) {
         return -1;
     }
@@ -2742,7 +2765,6 @@ int32_t common_speculative_on_target_batch(
         return -1;
     }
 
-    const float * first_hidden = hidden_rows_storage.data();
     const float * last_hidden = hidden_rows_storage.data() + (size_t) (batch.n_tokens - 1) * features.width;
     mtp_store_target_hidden(*mtp_state, seq_id, last_hidden, features.width);
 
@@ -2751,16 +2773,41 @@ int32_t common_speculative_on_target_batch(
         return 0;
     }
 
-    if (is_prompt_warmup) {
-        if (!llama_set_draft_input_hidden_state_copy(mtp_state->ctx_mtp, hidden_rows_storage.data(), hidden_rows_storage.size())) {
-            return -1;
-        }
-        const int32_t ret = mtp_update_kv_cache(mtp_state->ctx_mtp, batch, true);
-        mtp_invalidate_cached_draft(*mtp_state, seq_id);
-        return ret;
+    if (!is_prompt_warmup) {
+        return mtp_accept_batch(*mtp_state, batch, seq_id, hidden_rows_storage.data());
     }
 
-    return mtp_accept_batch(*mtp_state, batch, seq_id, first_hidden);
+    const bool uses_shifted_hidden_rows = mtp_model_uses_recurrent_conditioning(*mtp_state);
+    std::vector<float> previous_hidden_storage;
+    if (uses_shifted_hidden_rows) {
+        const auto hidden_it = mtp_state->target_hidden_by_seq.find(seq_id);
+        if (hidden_it != mtp_state->target_hidden_by_seq.end() && (int32_t) hidden_it->second.size() == features.width) {
+            previous_hidden_storage = hidden_it->second;
+        } else {
+            previous_hidden_storage.assign(features.width, 0.0f);
+        }
+    }
+
+    const float * conditioned_hidden_rows = hidden_rows_storage.data();
+    std::vector<float> conditioned_hidden_storage;
+    if (uses_shifted_hidden_rows) {
+        conditioned_hidden_storage.resize(hidden_rows_storage.size());
+        std::copy(previous_hidden_storage.begin(), previous_hidden_storage.end(), conditioned_hidden_storage.begin());
+        if (batch.n_tokens > 1) {
+            std::copy(
+                hidden_rows_storage.begin(),
+                hidden_rows_storage.begin() + (size_t) (batch.n_tokens - 1) * features.width,
+                conditioned_hidden_storage.begin() + features.width);
+        }
+        conditioned_hidden_rows = conditioned_hidden_storage.data();
+    }
+
+    if (!llama_set_draft_input_hidden_state_copy(mtp_state->ctx_mtp, conditioned_hidden_rows, hidden_rows_storage.size())) {
+        return -1;
+    }
+    const int32_t ret = mtp_update_kv_cache(mtp_state->ctx_mtp, batch, true);
+    mtp_invalidate_cached_draft(*mtp_state, seq_id);
+    return ret;
 }
 
 common_speculative_type common_speculative_current_type(const common_speculative * spec) {

include/llama.h

@@ -698,6 +698,8 @@ extern "C" {
 
     LLAMA_API bool llama_model_is_split_mode_graph(const struct llama_model * model);
 
+    LLAMA_API const char * llama_model_arch_string(const struct llama_model * model);
+
     // Returns 0 on success
     LLAMA_API uint32_t llama_model_quantize(
             const char * fname_inp,

src/llama-arch.cpp

@@ -1,5 +1,6 @@
 #include "llama-arch.h"
 #include "llama-impl.h"
+#include "llama-model.h"
 
 #include <map>
 
@@ -101,6 +102,15 @@ llm_arch llm_arch_from_string(const std::string & name) {
     return LLM_ARCH_UNKNOWN;
 }
 
+const char * llama_model_arch_string(const struct llama_model * model) {
+    static const char * unknown = "unknown";
+    if (!model) return unknown;
+    if (auto it = LLM_ARCH_NAMES.find(model->arch); it != LLM_ARCH_NAMES.end()) {
+        return it->second;
+    }
+    return unknown;
+}
+
 static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_GENERAL_TYPE,                  "general.type"                          },
     { LLM_KV_GENERAL_ARCHITECTURE,          "general.architecture"                  },