Hybrid recurrent cache #13979

gabe-l-hart · 2025-06-02T19:50:36Z

This is a re-opened version of #13904 after #13746 was merged

Description

This PR introduces the llama_kv_cache_hybrid_recurrent cache implementation. It follows the pattern of llama_kv_cache_unified_iswa by holding two child cache instances and implementing the interface logic such that it manages both correctly for the appropriate layers.

Changes

The main change in this PR is the addition of llama_kv_cache_hybrid_recurrent in llama-kv-cache-hybrid-recurrent.*. In addition to this, the PR does the following:

Add the llama_model_is_hybrid_recurrent public API (akin to llama_model_is_recurrent)
Add LLM_KV_ATTENTION_LAYER_INDICES as an hparam to hold the indices of the layers that should use attention (versus recurrent)
- This part is not well aligned with iswa, but that mechanism also isn't particularly extensible. It might be more appropriate to have a generic mechanism for indicating the type of caching to use for each layer, but that would start to approach the generic hybrid implementation that I originally attempted which ended up being too abstract (feat: Hybrid unified/recurrent cache #13276).
Abstracting utilities in llm_graph_context that need a specific type of cache to use getters (get_state_unified / get_state_recurrent) that will properly handle llama_kv_cache_hybrid_recurrent
Make n_embd_k_s / n_embd_v_s layer-dependent and use layer indices when calling them in the existing cache implementations
Add layer filtering to llama_kv_cache_recurrent
Updates the logic in llama_model::create_memory to use llm_arch_is_recurrent and llm_arch_is_hybrid_recurrent rather than relying on adding models to the switch statement which was redundant with the implementation of these functions

ggerganov · 2025-06-03T06:14:19Z

src/llama-graph.cpp

+const llama_kv_cache_unified_state * llm_graph_context::get_state_unified() const {
+    const auto * umstate = dynamic_cast<const llama_kv_cache_unified_state *>(mstate);
+    if (!umstate) {
+        const auto hmstate = dynamic_cast<const llama_kv_cache_hybrid_recurrent_state *>(mstate);
+        if (hmstate) {
+            umstate = hmstate->get_state_attn();
+        }
+    }
+    GGML_ASSERT(umstate);
+    return umstate;
+}
+
+const llama_kv_cache_recurrent_state * llm_graph_context::get_state_recurrent() const {
+    const auto * rmstate = dynamic_cast<const llama_kv_cache_recurrent_state *>(mstate);
+    if (!rmstate) {
+        const auto hmstate = dynamic_cast<const llama_kv_cache_hybrid_recurrent_state *>(mstate);
+        if (hmstate) {
+            rmstate = hmstate->get_state_recurrent();
+        }
+    }
+    GGML_ASSERT(rmstate);
+    return rmstate;
+}
+


These dynamic casts should not be necessary. Instead you need a new llm_graph_context::build_attn_inp_kv_hybrid_recurrent() method, similar to llm_graph_context::build_attn_inp_kv_unified_iswa().

I'm working through this now and a couple of questions are coming up:

Would it be best to combine build_inp_s_copy with build_attn_inp_kv for hybrid so that models call just one "build inputs" function, or keep them separate for simplicity?

For the build_attn methods, each has a corresponding llm_graph_input_attn_* class. The build_inp_s_* methods don't have this pattern which would make this a bit harder to have code reuse. Are there plans to refactor that further @compilade?

In the mamba2 branch, s_mask seems to be totally removed. I'd prefer not to do all of the boilerplate for duplicating build_inp_s_mask for the hybrid recurrent case if that's definitely going to be going away. Is there any reason that might stick around past the merge of mamba2?

Answering out of order, but it should still make sense:

2. For the build_attn methods, each has a corresponding llm_graph_input_attn_* class. The build_inp_s_* methods don't have this pattern

They do follow this pattern, see

llama.cpp/src/llama-graph.h

Line 190 in 7e00e60

class llm_graph_input_s_copy : public llm_graph_input_i {

(this is on the current master)

I think you might mean the build_attn_* methods also return instances of llm_graph_input_attn_*?
That seems to be directly related to llm_graph_context::build_attn() which has multiple implementations which differ by the type of the first argument (e.g. for llm_graph_input_attn_kv_unified, llm_graph_input_attn_no_cache, etc.)

Are there plans to refactor that further @compilade?

Not really, outside of removing s_mask (and related functions and classes) as part of #13834.

Would it be best to combine build_inp_s_copy with build_attn_inp_kv for hybrid so that models call just one "build inputs" function, or keep them separate for simplicity?

Personally, I think it would be simpler to keep them separate, because they are fundamentally different (one is intended to be used by build_copy_mask_state (renamed to build_recurrent_state in #13834), while the other is used by build_attn), and they are pretty much independent, even in hybrid models (at least for Jamba, the recurrent and self-attention layers are mostly independent on that front).

I don't see how build_attn would ever need s_copy.

build_inp_s_copy and build_inp_attn_kv_* are called once at the beginning of the graph, while build_attn and build_recurrent_state are called once per layer (where applicable, and so usually different layers for both).

3. Is there any reason [s_mask] might stick around past the merge of mamba2?

No reason to keep it, s_mask will be removed. Its functionality is redundant with s_copy, and otherwise prevents minimizing unnecessary state copies. It was used to clear the states, but the same can be done through inp_s_copy and clearing by copying a zero-ed state (which is the rs_z'th state in the mamba2 branch (and #13834)).

Thanks, that's super helpful! I was missing the distinction between build_attn_inp and build_attn which makes perfect sense.

Personally, I think it would be simpler to keep them separate

I agree on my personal gut feeling, so I'll go with this and see how it feels once complete.

Ok, I think this feels a lot cleaner now. For build_inp_s_copy, I opted to add an optional parameter so that the caller can take ownership of casting the cache state rather than duplicating the function into build_inp_s_copy_hybrid. That felt a little cleaner w.r.t. code reuse, but I'm happy to do a separate method if that's preferred.

It looks like there's one more place that will need changing in build_copy_mask_state (renamed to build_recurrent_state on mamba2). Similar to build_inp_s_copy, I think the cleanest way to do this for code reuse is to add an optional parameter that, if unset, will use the current logic of casting mstate.

compilade · 2025-06-02T20:07:00Z

src/llama-kv-cache-hybrid-recurrent.cpp

+        // TODO: will the recurrent cache be in an undefined state at this point?
+        LLAMA_LOG_ERROR("%s: failed to prepare recurrent ubatches\n", __func__);


Yes, but that will be fixed in #13834

(Noting here in case this gets merged first so that I don't forget to update the comment)

Should I strip out this TODO at this point?

src/llama-kv-cache-hybrid-recurrent.cpp

gabe-l-hart · 2025-06-09T17:02:28Z

@ggerganov I've noticed that the Server tests are consistently failing on this branch, but I see them passing on other PRs (eg #14081). The failures seem to be around retried connections to /slots/0?action=restore (here). I've noticed a couple of things when trying to repro locally:

For all of the server tests, despite the health check ping loop, I need to manually inject a sleep after server.start() to get any of the tests to pass on my M3 Max 64GB, otherwise the first endpoint call will almost always return an unexpected result (invalid json, or 5xx error). To me, this speaks to the server reporting /health success before it's logically "ready." This is probably correct in that the server is healthy enough to take requests, but the tests are treating this as a readiness check which would indicate that the server is fully initialized which seems incorrect (similar to liveness vs readiness in kubernetes). I wasn't able to find any open issues related to this, so I'd be curious if there's a discussion somewhere of readiness vs liveness?
Once I do insert the manual sleep in test_slot_save_restore, the test passes locally for me. This speaks to some difference between my local system and the GH Action runners. I'm not familiar enough with the server code to know whether this would be caused by some kind of memory constraint, disk speed constraint, etc, but my guess would be something like slot1.bin not being fully written somehow. This wouldn't explain why it passes on other branches though. Any insight on the nature of these tests would be much appreciated as I try to debug!

This is an attempt to handle race conditions between /health returning OK and the other endpoints not returning successfully. ggml-org#13979 (comment) Branch: HybridRecurrentCache Signed-off-by: Gabe Goodhart <[email protected]>

gabe-l-hart · 2025-06-09T17:38:16Z

I've tried adding retry logic for all requests in 39a93b3 to work around the race between /health, but I'm not sure if this is just going to mask an underling issue.

ggerganov · 2025-06-09T19:38:31Z

The changes to the server tests should not be needed. Let's revert the commit for now and I'll investigate.

gabe-l-hart · 2025-06-09T20:14:12Z

@ggerganov Thanks, it looks like those changes didn't fix the failures anyway, so definitely not the right fix. I've reset them out and can open an issue with details of what I see locally

gabe-l-hart · 2025-06-09T20:56:10Z

Issue for follow up on /health race condition with tests: #14092

gabe-l-hart · 2025-06-10T21:23:27Z

I've rebased on #13834. Drafting for now until it's merged

Also, split llama_model_is_recurrent into llm_arch_is_recurrent in llama-arch with llama_model_is_recurrent delegating to llm_arch_is_recurrent. The same split is done for hybird. This is needed because there are places where the llama_model has not yet been initialized but we need to check if the model is recurrent (specifically for the per-layer recurrent check array in hparams). Branch: GraniteFour Signed-off-by: Gabe Goodhart <[email protected]>

Branch: GraniteFour

…s in hparams Branch: GraniteFour Signed-off-by: Gabe Goodhart <[email protected]>

…l is recurrent Branch: GraniteFour Signed-off-by: Gabe Goodhart <[email protected]>

Branch: HybridCache Signed-off-by: Gabe Goodhart <[email protected]>

Branch: GraniteFour Signed-off-by: Gabe Goodhart <[email protected]>

This follows the pattern in iswa where the two child caches are held explicitly to support the case where a model requires a single attention cache and a single recurrent cache where each layer uses exactly one of the caches. This is a rewrite of the more generic approach in the original hybrid cache PR: ggml-org#13276 Branch: HybridRecurrentCache Signed-off-by: Gabe Goodhart <[email protected]>

This includes a refactor of the create_memory logic to avoid needing to use the arch enum explicitly unless a model needs explicit cache instantiation logic beyond the standard logic for recurrent, hybrid, unified, and iswa. Branch: HybridRecurrentCache Signed-off-by: Gabe Goodhart <[email protected]>

Branch: HybridRecurrentCache Signed-off-by: Gabe Goodhart <[email protected]>

NOTE: I intentionally did not add support for s_mask since it will be going away soon Branch: HybridRecurrentCache Signed-off-by: Gabe Goodhart <[email protected]>

Branch: GraniteFour Signed-off-by: Gabe Goodhart <[email protected]>

…lits in unified cache Branch: GraniteFour Signed-off-by: Gabe Goodhart <[email protected]>

…he interface Branch: HybridRecurrentCache Signed-off-by: Gabe Goodhart <[email protected]>

Branch: GraniteFour Signed-off-by: Gabe Goodhart <[email protected]>

Branch: HybridRecurrentCache Signed-off-by: Gabe Goodhart <[email protected]>

gabe-l-hart · 2025-06-10T22:22:35Z

That was quick! Undrafting now that #13834 is merged

gabe-l-hart · 2025-06-10T22:24:39Z

src/llama-graph.h

@@ -242,7 +243,7 @@ class llm_graph_input_attn_kv_unified : public llm_graph_input_i {
        cparams(cparams),
        kv_state(kv_state) {
    }
-    ~llm_graph_input_attn_kv_unified() = default;
+    virtual ~llm_graph_input_attn_kv_unified() = default;


I don't think this virtual is needed. I think I added it when I was attempting to have the hybrid input inherit from the unified input

…empt Branch: HybridRecurrentCache Signed-off-by: Gabe Goodhart <[email protected]>

Branch: HybridRecurrentCache Signed-off-by: Gabe Goodhart <[email protected]>

gabe-l-hart · 2025-06-10T22:31:18Z

src/llama-hparams.h

@@ -181,10 +184,13 @@ struct llama_hparams {

    // dimension of the rolling state embeddings
    // corresponds to Mamba's conv_states size or RWKV's token_shift states size
-    uint32_t n_embd_k_s() const;
+    uint32_t n_embd_k_s(uint32_t il = 0) const;


I initially added these with the default value of 0 to avoid needing to change code that had no chance of being in the path of a hybrid model. After this PR, I think the only place that's true is the various rwkv methods in llama-graph and llama-model. Would it be better to always require an argument?

gabe-l-hart · 2025-06-10T22:34:00Z

src/llama-graph.h

+                                     int32_t   state_size,
+                                     int32_t   n_seqs,
+                                        bool   avoid_copies = false,
+        const llama_kv_cache_recurrent_state * kv_state = nullptr) const;


NIT: This was the closest I could come up to for consistent indentation, but it's... ugly? Very open to preferred style suggestions.

Non-NIT: I don't love adding this, especially after avoid_copies since it now forces all layer builders that may be in a hybrid architecture to explicitly set avoid_copies.

gabe-l-hart · 2025-06-10T22:37:33Z

src/llama-kv-cache-hybrid-recurrent.cpp

+}
+
+llama_pos llama_kv_cache_hybrid_recurrent::seq_pos_min(llama_seq_id seq_id) const {
+    // the min of the total cache is the max of the two caches' min values


I'd definitely love a double-check on this logic!

gabe-l-hart · 2025-06-10T22:40:56Z

src/llama-kv-cache-hybrid-recurrent.h

+    // llama_kv_cache_hybrid_recurrent specific API
+    //
+
+    llama_kv_cache_unified   * get_kv_attn     () const;


I don't think I've ever actually used the friend keyword in real code, but I think we could make llama_kv_cache_hybrid_recurrent_state a friend and then make these private for slightly better encapsulation.

Actually, if we made it a friend, these methods could just be removed since the child unique_ptr instances could just be accessed directly.

gabe-l-hart · 2025-06-10T22:46:56Z

src/llama-kv-cache-hybrid-recurrent.h

+    //
+
+    llama_memory_state_ptr init_batch(
+            const llama_batch & batch,


NIT to self: This doesn't follow other indentation conventions

gabe-l-hart · 2025-06-10T22:49:24Z

src/llama-kv-cache-recurrent.cpp

@@ -88,8 +94,8 @@ llama_kv_cache_recurrent::llama_kv_cache_recurrent(
        ggml_tensor * v = ggml_new_tensor_1d(ctx, type_v, n_embd_v_gqa*kv_size);
        ggml_format_name(k, "cache_k_l%d", i);
        ggml_format_name(v, "cache_v_l%d", i);
-        k_l.push_back(k);


This push_back was broken with the addition of the filter above to skip non-recurrent layers.

gabe-l-hart · 2025-06-10T22:51:00Z

src/llama-kv-cache-recurrent.cpp

-                 uint32_t   kv_size,
-                 uint32_t   n_seq_max) : hparams(model.hparams), n_seq_max(n_seq_max) {
+        const llama_model &  model,
+          layer_filter_cb && filter,


I opted to use an opaque filter here rather than explicitly using the model's hparams.recurrent_layer() method for better separation, but it's still not totally clean since n_embd_k/v_s are used below based on recurrent_layer(), so there could be a potential mismatch if the filter included some layers that then had n_embd_k/v_s return 0.

gabe-l-hart · 2025-06-10T22:52:13Z

src/llama-kv-cache-recurrent.h

@@ -15,13 +15,18 @@
 //       see the implementation of llama_kv_cache_unified_state_i for an example how to do it
 class llama_kv_cache_recurrent : public llama_memory_i {
 public:
+
+    // this callback is used to filter out layers that should not be included in the cache


This is copy-pasta from the unified cache. I had originally hoisted it to the llama_kv_cache intermediate base class, but since that's been removed, I figured it was better to just copy it rather than put it into the top-level interface class.

gabe-l-hart · 2025-06-10T22:54:52Z

src/llama-model.cpp

@@ -469,6 +470,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
    std::fill(hparams.n_head_arr.begin(),    hparams.n_head_arr.end(),    0);
    std::fill(hparams.n_head_kv_arr.begin(), hparams.n_head_kv_arr.end(), 0);
    std::fill(hparams.n_ff_arr.begin(),      hparams.n_ff_arr.end(),      0);
+    std::fill(


On this branch, this is somewhat incomplete. For the Granite 4 branch, I have this array further overridden by parsing LLM_KV_ATTENTION_LAYER_INDICES. I'm wondering if that logic should be universal enough that I should hoist it to this portion here to be done in a model-agnostic way.

gabe-l-hart · 2025-06-10T22:56:12Z

src/llama-model.cpp

@@ -13198,6 +13203,8 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
    llama_memory_i * res;

    switch (arch) {
+        // Models that need specific instantiation should be handled in the


This is a slight architectural refactor to avoid the need to duplicate the arch switch statement between this and llm_arch_is_[hybrid]_recurrent. I'm not tied to this, but it feels a little cleaner to hold a single source-of-truth for determining which models are hybrid/hybrid-recurrent.

gabe-l-hart · 2025-06-10T22:58:35Z

@compilade @ggerganov I think this one is ready for a full review now based on the list shared on the earlier refactor PR

gabe-l-hart mentioned this pull request Jun 2, 2025

Hybrid recurrent cache #13904

Closed

1 task

ggerganov reviewed Jun 3, 2025

View reviewed changes

compilade reviewed Jun 3, 2025

View reviewed changes

gabe-l-hart force-pushed the HybridRecurrentCache branch 7 times, most recently from ab918bb to 60ca3ba Compare June 9, 2025 16:06

gabe-l-hart requested a review from ngxson as a code owner June 9, 2025 17:36

github-actions bot added examples python python script changes server labels Jun 9, 2025

gabe-l-hart force-pushed the HybridRecurrentCache branch from 39a93b3 to 60ca3ba Compare June 9, 2025 20:13

gabe-l-hart mentioned this pull request Jun 9, 2025

Misc. bug: Server tests /health race conditions #14092

Open

gabe-l-hart force-pushed the HybridRecurrentCache branch 2 times, most recently from 7958d84 to 3669876 Compare June 10, 2025 21:22

gabe-l-hart marked this pull request as draft June 10, 2025 21:22

gabe-l-hart added 4 commits June 10, 2025 16:21

feat: Add c++ side constants for attention layer indices hparam

25ebc6b

Branch: GraniteFour

feat: Add support for distinguishing recurrent vs non-recurrent layer…

a1829a7

…s in hparams Branch: GraniteFour Signed-off-by: Gabe Goodhart <[email protected]>

feat: Auto-fill hparams.recurrent_layer_arr based on whether the mode…

683ba84

…l is recurrent Branch: GraniteFour Signed-off-by: Gabe Goodhart <[email protected]>

gabe-l-hart added 13 commits June 10, 2025 16:21

feat: Add layer filter to recurrent cache

eed7811

Branch: HybridCache Signed-off-by: Gabe Goodhart <[email protected]>

fix: Use per-layer sizing everywhere in kv caches

5d0f877

Branch: GraniteFour Signed-off-by: Gabe Goodhart <[email protected]>

fix: Fix wrong bool condition for split equal in hybrid cache

d224255

Branch: HybridRecurrentCache Signed-off-by: Gabe Goodhart <[email protected]>

fix: Fix shift logic to defer to unified cache

71ce48e

Branch: HybridRecurrentCache Signed-off-by: Gabe Goodhart <[email protected]>

feat: Support hybrid recurrent in llama-graph

8bac844

NOTE: I intentionally did not add support for s_mask since it will be going away soon Branch: HybridRecurrentCache Signed-off-by: Gabe Goodhart <[email protected]>

fix: Fix logic for initializing inputs and attn layers for hybrid caches

be2ba0f

Branch: GraniteFour Signed-off-by: Gabe Goodhart <[email protected]>

fix: Use @compilade's suggested fix for seq_id indexing with equal sp…

4988950

…lits in unified cache Branch: GraniteFour Signed-off-by: Gabe Goodhart <[email protected]>

fix: Update recurrent cache for changes to remove intermediate kv_cac…

19bf5b2

…he interface Branch: HybridRecurrentCache Signed-off-by: Gabe Goodhart <[email protected]>

fix: Fix status for init_update sig for recurrent cache state

087cbab

Branch: GraniteFour Signed-off-by: Gabe Goodhart <[email protected]>

fix: Add missing padding to n_ctx for hybrid cache construction

2ed5a45

Branch: GraniteFour Signed-off-by: Gabe Goodhart <[email protected]>

fix: Update clear signature for data argument after rebase

8c59841

Branch: HybridRecurrentCache Signed-off-by: Gabe Goodhart <[email protected]>

gabe-l-hart force-pushed the HybridRecurrentCache branch from 3669876 to 8c59841 Compare June 10, 2025 22:22

gabe-l-hart marked this pull request as ready for review June 10, 2025 22:22

gabe-l-hart commented Jun 10, 2025

View reviewed changes

gabe-l-hart added 2 commits June 10, 2025 16:26

fix: Remove errant virtual destructor leftover from previous impl att…

c831e76

…empt Branch: HybridRecurrentCache Signed-off-by: Gabe Goodhart <[email protected]>

fix: Use per-layer n_embd_k/v_s calls for mamba (1) layers

bb87dbf

Branch: HybridRecurrentCache Signed-off-by: Gabe Goodhart <[email protected]>

gabe-l-hart commented Jun 10, 2025

View reviewed changes

		// TODO: will the recurrent cache be in an undefined state at this point?
		LLAMA_LOG_ERROR("%s: failed to prepare recurrent ubatches\n", __func__);

Hybrid recurrent cache #13979

Are you sure you want to change the base?

Hybrid recurrent cache #13979

Conversation

gabe-l-hart commented Jun 2, 2025

Description

Changes

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

compilade Jun 4, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

gabe-l-hart commented Jun 9, 2025

Uh oh!

gabe-l-hart commented Jun 9, 2025

Uh oh!

ggerganov commented Jun 9, 2025

Uh oh!

gabe-l-hart commented Jun 9, 2025

Uh oh!

gabe-l-hart commented Jun 9, 2025

Uh oh!

gabe-l-hart commented Jun 10, 2025

Uh oh!

gabe-l-hart commented Jun 10, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

gabe-l-hart Jun 10, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

gabe-l-hart commented Jun 10, 2025

Uh oh!

Uh oh!

compilade Jun 4, 2025 •

edited

Loading

gabe-l-hart Jun 10, 2025 •

edited

Loading