Hybrid recurrent cache

[gabe-l-hart]

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]

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().

[gabe-l-hart]

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

1. 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?
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 which would make this a bit harder to have code reuse. Are there plans to refactor that further @compilade?
3. 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?

[compilade]

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.

1. 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)).

[gabe-l-hart]

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.

[gabe-l-hart]

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.

[gabe-l-hart]

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]

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)

[gabe-l-hart]

Should I strip out this TODO at this point?

[gabe-l-hart]

@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:

1. 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?

2. 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!

[gabe-l-hart]

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]

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

[gabe-l-hart]

@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]

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

[gabe-l-hart]

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

[gabe-l-hart]

That was quick! Undrafting now that #13834 is merged

[gabe-l-hart]

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

[younesbelkada]

Hi @gabe-l-hart - thanks for your huge work. Chiming in this PR to ask a question: how do you think we should handle cache formats where for every layer we have both the attention and SSM states ? This design seems to be tailored for architectures that have either SSM or Attention for each decoder block

[gabe-l-hart]

Hi @younesbelkada! That sounds like a really interesting architecture. Can you point to any papers or draft models that use this approach? I'd be really curious to check them out. The main advantage of the hybrid recurrent architecture where each layer is either SSM or Attention (but not both) is that you get the cache size scaling advantages of SSMs for most layers without sacrificing the ability to attend to all tokens for the whole model.

My first pass at a standalone hybrid cache (#13276) attempted to be a more generic representation of "each layer can have its own cache type" where the parent cache could have any number of children. This ran into a number of challenges during the recent interface refactoring, and ultimately was incompatible with some of the work to encapsulate the interactions with the cache, so at Georgi's suggestion, I pivoted to this less-abstract implementation.

Following the pattern established by iswa and now hybrid_recurrent, I would think implementing such a cache would look like:

1. Creating another new subclass of llama_memory_i that owns the child caches (similar to hybrid_recurrent)
   • The various interfaces would probably be implemented with "AND" semantics instead of "OR" semantics for child layers
2. Creating the corresponding subclass of llama_memory_state_i
3. Use a different set of layer filters when constructing the parent cache with the children such that the children apply to the right layers (which are now overlapping sets)

The place where I see the most incompatibility is in the hparams.recurrent_layer() method which right now is a "yes/no" and if true implies ! attention_layer (explicitly in the n_embd_k_s / n_embd_v_s methods). One possible solution here would be for the hparams to hold some kind of get_cache_type(il) -> set{CACHE_TYPE} method that would return a set of cache type enum values for a given layer index. This might also be able to replace the llm_arch_is_recurrent / llm_arch_is_hybrid_recurrent interfaces.

@compilade @ggerganov do you foresee the need to allow a single layer to use multiple cache types as something we'll want to support in the future? If so, we might want to implement that more-generic interface in hparams now to avoid an API breaking change down the road.

[younesbelkada]

Thank you @gabe-l-hart for your detailed response 🙏
Currently there is an effort to try to integrate Falcon-H1 models into main llama.cpp: https://huggingface.co/collections/tiiuae/falcon-h1-6819f2795bc406da60fab8df (for more details you can check the technical blogpost here, this architecture employs a parallel hybrid design where you have both SSM and Attention on each decoder block) - we have a public fork of llama.cpp that we are currently maintaining here: https://github.com/tiiuae/llama.cpp-Falcon-H1 but we are refactoring the changes we made so that it will be compatible with the changes on master (including this PR and #9126) and merge the changes in llama.cpp master (related issue: #13681)
Right now we tried to build on top of this PR and the Mamba2 PR from @compilade and faced this issue so we are not sure what is the approach to take. What you suggested sounds good - we will be happy to work on whatever you think is the best suggestion!

[gabe-l-hart]

Ah! I actually think someone pointed me at your fork a week ago and I lost it in my exploded number of tabs. I'll definitely check it out and see if there are any places this should tweak to accommodate the Falcon-H1 arch.

[gabe-l-hart]

For context, my motivation is very similar to yours. I'm trying to get Granite 4 support added which is also a combo of mamba2 and attention (GraniteMoeShared for the attention layers). My draft PR is #13550, but my most recent attempt to merge master and mamba2 failed pretty badly, so I'm still debugging why (and I'm guessing compilade is also working on doing it the right way).

[gabe-l-hart]

Yes, probably we have to introduce the notion of has_recurrent_state(il), has_attention(il) and these can be both true in such cases.

This makes sense. I'll give this a shot soon.

Of these uses, I think the two that would be the most problematic are n_embd_k/v_s since they're currently implemented using hparams.recurrent_layer and hparams.n_embd since that is used by both the unified and recurrent caches and I think might need to have different values for the two if they were being shared by the same layer.

@ggerganov with the removal of n_embd_k/v_s layer indexing, I think that just leaves hparams.n_embd. Do you have a quick thought on whether this will need to be split between attn/recurrent portions for a single layer?

[ggerganov]

Does H1 actually need different hparams.n_embd for the different layers? The hparams.n_embd is the size of the hidden state that goes through the layers. I haven't looked at the links, but I doubt this parameter would be vary with layers.

[gabe-l-hart]

I'm not clear on that (haven't unboxed it far enough either). @younesbelkada I haven't dug into your fork deeply yet, but I'm guessing you may have had to tackle this question and might know which hparams would need to be different between the attention / recurrent portions of the cache for a given layer.

[gabe-l-hart]

I've added the ability to specify custom filters in the constructor to llama_kv_cache_hybrid_recurrent which will default to checking hparams.recurrent_layer if not set. This should allow H1 to implement its own "always on" filter for all layers and then add a clause to create_memory in the model-specific section to construct the hybrid cache with that filter.

[younesbelkada]

Thank you @gabe-l-hart @ggerganov !
If I understand correctly from the last comment from @gabe-l-hart, we can already give it a shot for H1 now. If you confirm, we'll start from your synced branch you shared on the Mamba2 PR

Does H1 actually need different hparams.n_embd for the different layers?

I can confirm we actually use the same n_embed across all layers since we aggregate the hidden states of Attention and SSM mixers by summing them - however the KV cache will have different shapes

-> SSM cache shape: https://github.com/huggingface/transformers/blob/main/src/transformers/models/falcon_h1/modeling_falcon_h1.py#L100-L117
-> KV cache shape: head_dim x num_kv_heads

[ggerganov]

Aren't n_embd_k_gqa and n_embd_v_gqa equal to 0 for recurrent models? If yes, you can remove them from the recurrent implementation.

[gabe-l-hart]

I think that sounds right. It looks like the value here is ultimately determined by hparams.n_head_kv_arr which by default is populated as a copy of hparams.n_head_arr from %s.attention.head_count. For mamba, I see mamba.attention.head_count set to 0 and I suspect this would be true of all other recurrent models, but I'm not sure how to fully verify. I'll try removing them and see if the recurrent models I have work as expected, but I'll defer to @compilade on whether this is an "always" thing or a "just mamba* models" thing.

[ggerganov]

I feel like the naming of the recurrent cache is not great and it is causing some confusion. In a follow-up PR we have to do the following naming changes

• Rename llama_kv_cache_recurrent -> llama_recurrent_state
• Rename llama_kv_cache_hybrid_recurrent -> llama_memory_hybrid

This should be a good first step, although not ideal because "hybrid" will implicitly refer to "unified KV cache + recurrent state". For the future we have to figure out a naming scheme that would be able to differentiate between different types of hybrid memories. But this should be good enough for now.

[ggerganov]

This method is analogous to the build_attn_inp_ methods, so we have to model it in a similar way.

Replace this method with:

    // similar to build_attn_inp_kv_unified()
    llm_graph_input_rs * build_rs_inp() const;

Introduce new input class:

// similar to llm_graph_input_attn_kv_unified
// put the `s_copy` tensor in this class (similar to `kq_mask`)
class llm_graph_input_rs : public llm_graph_input_i;

In the future, this input class could be extended with additional input tensors that are needed by the recurrent cache if necessary (similar to the attention input classes).

Replace build_recurrent_state() and build_rwkv_shift_load() with overloads:

    // similar to build_attn()
    ggml_tensor * build_rs(
        llm_graph_input_rs * inp,
             ggml_cgraph * gf,
             ggml_tensor * s,             
                 int32_t   state_size,
                 int32_t   n_seqs,
                    bool   avoid_copies = false) const;

    // similar to build_attn()
    ggml_tensor * build_rwkv_token_shift_load(
        llm_graph_input_rs * inp,
             ggml_cgraph * gf,
      const llama_ubatch & ubatch,
                     int   il) const;

[ggerganov]

When this change is applied, we have to do a similar addition for the hybrid implementation. The basic pattern is that you need to introduce a new input class similar to llm_graph_input_attn_kv_unified and llm_graph_input_rs, but this one will contain inputs for both the attention and for the recurrent state.

So probably something like:

// this input class will have both the input tensors needed for the attention and for
// the recurrent state. see llm_graph_input_attn_kv_unified_iswa for example
class llm_graph_input_mem_hybrid : public llm_graph_input_i;

We then add overloads for build_attn(), build_rs() and build_rwkv_token_shift_load().

[ggerganov]

This change is not needed. We have to aim to not mix the recurrent logic with the attention logic.

[ggerganov]

These can be changed like this, similar to the iswa state:

Suggested change

	const llama_kv_cache_unified_state_ptr state_attn;
	const llama_kv_cache_recurrent_state_ptr state_recurrent;
	llama_memory_state_ptr state_attn;
	llama_memory_state_ptr state_recurrent;

[ggerganov]

For now let's hold-off adding this API. I don't think the llama_model_is_recurrent should exist either. The idea is that the user app should not have to know what is the underlying memory mechanism.