feat: Add can_seq_rm API to llama_kv_cache API

This will be key for the hybrid cache which needs to be able to validate
that all children can perform seq_rm cleanly before attempting to remove
the seq from any single child to avoid ending up in a corrupted state.

Branch: HybridCache

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

Author: gabe-l-hart

src/llama-kv-cache.cpp

@@ -463,6 +463,14 @@ void llama_kv_cache_unified::set_full() {
     head = 0;
 }
 
+bool llama_kv_cache_unified::can_seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) const {
+    GGML_UNUSED(seq_id);
+    GGML_UNUSED(p0);
+    GGML_UNUSED(p1);
+    // Unified attention cache can always do a sequence removal
+    return true;
+}
+
 llama_sbatch llama_kv_cache_unified::sbatch_init(const llama_batch & batch, bool logits_all) {
     return llama_sbatch(batch, hparams.n_embd, true, logits_all);
 }
@@ -1773,6 +1781,15 @@ void llama_kv_cache_unified_iswa::set_full() {
     kv_swa ->set_full();
 }
 
+bool llama_kv_cache_unified_iswa::can_seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) const {
+    GGML_UNUSED(seq_id);
+    GGML_UNUSED(p0);
+    GGML_UNUSED(p1);
+    // Unified attention caches can always do a sequence removal, so since both
+    // children can, the parent can as well.
+    return true;
+}
+
 llama_sbatch llama_kv_cache_unified_iswa::sbatch_init(const llama_batch & batch, bool logits_all) {
     pending.clear();
 
@@ -1968,39 +1985,33 @@ void llama_kv_cache_recurrent::clear() {
 }
 
 bool llama_kv_cache_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
-    uint32_t new_head = size;
+    if (!can_seq_rm(seq_id, p0, p1)) {
+        // could be fatal
+        return false;
+    }
 
+    uint32_t new_head = size;
     if (p0 < 0) {
         p0 = 0;
     }
-
     if (p1 < 0) {
         p1 = std::numeric_limits<llama_pos>::max();
     }
 
-    // models like Mamba or RWKV can't have a state partially erased
-    if (seq_id >= (int64_t) size) {
-        // could be fatal
-        return false;
-    }
     if (0 <= seq_id) {
         int32_t & tail_id = cells[seq_id].tail;
         if (tail_id >= 0) {
             const kv_cell & cell = cells[tail_id];
-            // partial intersection is invalid
-            if ((0 < p0 && p0 <= cell.pos) || (0 < p1 && p1 <= cell.pos)) {
-                return false;
-            }
+            // already validated in can_seq_rm
+            GGML_ASSERT(!((0 < p0 && p0 <= cell.pos) || (0 < p1 && p1 <= cell.pos)));
             // invalidate tails which will be cleared
             if (p0 <= cell.pos && cell.pos < p1) {
                 tail_id = -1;
             }
         }
     } else {
-        // seq_id is negative, then the range should include everything or nothing
-        if (p0 != p1 && (p0 != 0 || p1 != std::numeric_limits<llama_pos>::max())) {
-            return false;
-        }
+        // already validated in can_seq_rm
+        GGML_ASSERT(!(p0 != p1 && (p0 != 0 || p1 != std::numeric_limits<llama_pos>::max())));
     }
 
     for (uint32_t i = 0; i < size; ++i) {
@@ -2217,6 +2228,34 @@ void llama_kv_cache_recurrent::set_full() {
     n = size;
     head = 0;
 }
+bool llama_kv_cache_recurrent::can_seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) const {
+    if (p0 < 0) {
+        p0 = 0;
+    }
+
+    if (p1 < 0) {
+        p1 = std::numeric_limits<llama_pos>::max();
+    }
+    // models like Mamba or RWKV can't have a state partially erased
+    if (seq_id >= (int64_t) size) {
+        // could be fatal
+        return false;
+    }
+    if (0 <= seq_id) {
+        const int32_t & tail_id = cells[seq_id].tail;
+        if (tail_id >= 0) {
+            const kv_cell & cell = cells[tail_id];
+            // partial intersection is invalid
+            if ((0 < p0 && p0 <= cell.pos) || (0 < p1 && p1 <= cell.pos)) {
+                return false;
+            }
+        }
+    // seq_id is negative, then the range should include everything or nothing
+    } else if (p0 != p1 && (p0 != 0 || p1 != std::numeric_limits<llama_pos>::max())) {
+        return false;
+    }
+    return true;
+}
 
 llama_sbatch llama_kv_cache_recurrent::sbatch_init(
         const llama_batch & batch,

src/llama-kv-cache.h

@@ -37,6 +37,11 @@ struct llama_kv_cache : public llama_memory_i {
     // simulate full cache, used for allocating worst-case compute buffers
     virtual void set_full() = 0;
 
+    // sometimes it is useful to check whether a cache can remove a sequence
+    // before attempting to mutate the cache (eg a hybrid cache with multiple
+    // children to keep in sync)
+    virtual bool can_seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) const = 0;
+
     //
     // batch processing
     //
@@ -142,6 +147,8 @@ class llama_kv_cache_unified : public llama_kv_cache {
 
     void set_full() override;
 
+    bool can_seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) const override;
+
     llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) override;
     llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const override;
 
@@ -353,6 +360,8 @@ class llama_kv_cache_unified_iswa : public llama_kv_cache {
 
     void set_full() override;
 
+    bool can_seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) const override;
+
     llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) override;
     llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const override;
 
@@ -464,6 +473,8 @@ class llama_kv_cache_recurrent : public llama_kv_cache {
 
     void set_full() override;
 
+    bool can_seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) const override;
+
     llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) override;
     llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const override;