Merge remote-tracking branch 'origin/compilade/mamba2' into BambaAbstractMemory

This is definitely a surgical merge, so there are likely pieces still
missing, especially from llama-model.cpp

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

* origin/compilade/mamba2:
convert : fix flake8 lint
ggml : avoid multiply by D in GGML_OP_SSM_SCAN
ggml : remove unused fast broadcast path in GGML_MUL
metal : fix wrong number of tokens per sequence in SSM_SCAN
metal : fix SSM_SCAN state head offset
metal : add back n_seqs to SSM_SCAN args
metal : remove unused arguments for SSM_SCAN
metal : use log and exp instead of log1pf and expf in SSM_SCAN
metal : fix SSM_SCAN pipeline scope
metal : attempt to adapt SSM_SCAN for Mamba-2
llama : avoid redundant state copy for Mamba 1 and 2
convert_hf : prefer SentencePiece tokenizer for Mamba-2 when present
llama : add missing break
llama : remove unused variable
llama : fix Mamba-2 conv state saving
llama : support running Mamba-Codestral-7B-v0.1
ggml : SIMD ggml_ssm_scan for Mamba-2
llama : initial Mamba-2 support

Author: gabe-l-hart

convert_hf_to_gguf.py

@@ -4171,6 +4171,91 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
         return [(new_name, data_torch)]
 
 
+@ModelBase.register("Mamba2ForCausalLM")
+class Mamba2Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.MAMBA2
+
+    def set_vocab(self):
+        vocab_size = self.hparams["vocab_size"]
+        # Round vocab size to next multiple of 16
+        pad_vocab = self.hparams.get("pad_vocab_size_multiple", 16)
+        # pad using ceiling division
+        # ref: https://stackoverflow.com/a/17511341/22827863
+        vocab_size = -(vocab_size // -pad_vocab) * pad_vocab
+        self.hparams["vocab_size"] = vocab_size
+
+        if (self.dir_model / "tokenizer.model").is_file():
+            self._set_vocab_sentencepiece()
+        elif (self.dir_model / "tokenizer.model.v3").is_file():
+            # mamba-codestral
+            raise NotImplementedError(f"Please rename {self.dir_model / 'tokenizer.model.v3'} to {self.dir_model / 'tokenizer.model'}")
+        elif (self.dir_model / "tokenizer.json").is_file():
+            self._set_vocab_gpt2()
+        else:
+            # Use the GPT-NeoX tokenizer when no tokenizer files are present
+            self._set_vocab_builtin("gpt-neox", vocab_size)
+
+    def set_gguf_parameters(self):
+        d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
+        d_conv  = self.find_hparam(["conv_kernel",       "d_conv"],  optional=True) or 4
+        d_inner = self.find_hparam(["intermediate_size", "d_inner"], optional=True) or 2 * d_model
+        d_state = self.find_hparam(["state_size",        "d_state"], optional=True) or 128
+        head_dim = self.find_hparam(["head_dim"],                    optional=True) or 64
+        n_group = self.find_hparam(["n_groups"],                     optional=True) or 1
+
+        rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5
+
+        # Fail early for models which don't have a block expansion factor of 2
+        # TODO: does this really matter?
+        assert d_inner == 2 * d_model
+        assert d_inner % head_dim == 0
+
+        self.gguf_writer.add_context_length(2**20)  # arbitrary value; for those who use the default
+        self.gguf_writer.add_embedding_length(d_model)
+        self.gguf_writer.add_feed_forward_length(0)  # unused, but seemingly required when loading
+        self.gguf_writer.add_head_count(0)  # unused, but seemingly required when loading
+        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_ssm_conv_kernel(d_conv)
+        self.gguf_writer.add_ssm_inner_size(d_inner)
+        self.gguf_writer.add_ssm_state_size(d_state)
+        self.gguf_writer.add_ssm_time_step_rank(d_inner // head_dim)
+        self.gguf_writer.add_ssm_group_count(n_group)
+        self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+
+        if name.startswith("model.backbone") or name.startswith("model.lm_head"):
+            # map Mamba-Codestral-7B-v0.1 tensor names to the names used by Mamba-2
+            name = name.removeprefix("model.")
+
+        if name.endswith(".dt_bias"):
+            name = name.rpartition(".dt_bias")[0] + ".dt_proj.bias"
+
+        new_name = self.map_tensor_name(name)
+
+        if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.SSM_CONV1D, bid):
+            data_torch = data_torch.squeeze()
+        elif any(self.match_model_tensor_name(new_name, t, bid, suffix="") for t in [
+            gguf.MODEL_TENSOR.SSM_A,
+            gguf.MODEL_TENSOR.SSM_D,
+        ]):
+            # unsqueeze A to use similar shape semantics as Mamba-1
+            # (D is also unsqueezed, but for more straightforward broadcast internally)
+            data_torch = data_torch.reshape((*data_torch.shape, 1))
+        elif self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.SSM_NORM, bid):
+            d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
+            d_inner = self.find_hparam(["intermediate_size", "d_inner"], optional=True) or 2 * d_model
+            n_group = self.hparams.get("n_groups", 1)
+            data_torch = data_torch.reshape((n_group, d_inner // n_group))
+
+        if name.endswith(".A_log"):
+            logger.debug("A_log --> A ==> " + new_name)
+            data_torch = -torch.exp(data_torch)
+
+        yield (new_name, data_torch)
+
+
 @ModelBase.register("CohereForCausalLM")
 class CommandR2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.COMMAND_R

ggml/include/ggml.h

@@ -1854,7 +1854,8 @@ extern "C" {
             struct ggml_tensor  * dt,
             struct ggml_tensor  * A,
             struct ggml_tensor  * B,
-            struct ggml_tensor  * C);
+            struct ggml_tensor  * C,
+            struct ggml_tensor  * ids);
 
     // partition into non-overlapping windows with padding if needed
     // example:

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -463,26 +463,25 @@ typedef struct {
 typedef struct {
     int64_t  d_state;
     int64_t  d_inner;
+    int64_t  n_head;
+    int64_t  n_group;
     int64_t  n_seq_tokens;
     int64_t  n_seqs;
-    uint64_t nb00;
     uint64_t nb01;
     uint64_t nb02;
-    uint64_t nb10;
+    uint64_t nb03;
     uint64_t nb11;
     uint64_t nb12;
     uint64_t nb13;
-    uint64_t nb20;
     uint64_t nb21;
     uint64_t nb22;
-    uint64_t nb30;
     uint64_t nb31;
-    uint64_t nb40;
     uint64_t nb41;
     uint64_t nb42;
-    uint64_t nb50;
+    uint64_t nb43;
     uint64_t nb51;
     uint64_t nb52;
+    uint64_t nb53;
 } ggml_metal_kargs_ssm_scan;
 
 typedef struct {

ggml/src/ggml-metal/ggml-metal.m

@@ -189,6 +189,7 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
     GGML_METAL_KERNEL_TYPE_NORM,
     GGML_METAL_KERNEL_TYPE_SSM_CONV_F32,
     GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32,
+    GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32_GROUP,
     GGML_METAL_KERNEL_TYPE_RWKV_WKV6_F32,
     GGML_METAL_KERNEL_TYPE_RWKV_WKV7_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,
@@ -1125,6 +1126,7 @@ @implementation GGMLMetalClass
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM,                            norm,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_CONV_F32,                    ssm_conv_f32,                    true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32,                    ssm_scan_f32,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32_GROUP,              ssm_scan_f32_group,              true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RWKV_WKV6_F32,                   rwkv_wkv6_f32,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RWKV_WKV7_F32,                   rwkv_wkv7_f32,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,                  mul_mv_f32_f32,                  has_simdgroup_reduction);
@@ -2586,71 +2588,91 @@ static bool ggml_metal_encode_node(
                 struct ggml_tensor * src3 = node->src[3];
                 struct ggml_tensor * src4 = node->src[4];
                 struct ggml_tensor * src5 = node->src[5];
+                struct ggml_tensor * src6 = node->src[6];
 
                 GGML_ASSERT(src3);
                 GGML_ASSERT(src4);
                 GGML_ASSERT(src5);
+                GGML_ASSERT(src6);
 
                 size_t offs_src3 = 0;
                 size_t offs_src4 = 0;
                 size_t offs_src5 = 0;
+                size_t offs_src6 = 0;
 
                 id<MTLBuffer> id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil;
                 id<MTLBuffer> id_src4 = src4 ? ggml_metal_get_buffer(src4, &offs_src4) : nil;
                 id<MTLBuffer> id_src5 = src5 ? ggml_metal_get_buffer(src5, &offs_src5) : nil;
+                id<MTLBuffer> id_src6 = src6 ? ggml_metal_get_buffer(src6, &offs_src6) : nil;
 
-                const int64_t  ne30 = src3->ne[0]; GGML_UNUSED(ne30);
+                const int64_t  ne30 = src3->ne[0];
                 const int64_t  ne31 = src3->ne[1]; GGML_UNUSED(ne31);
 
-                const uint64_t nb30 = src3->nb[0];
+                const uint64_t nb30 = src3->nb[0]; GGML_UNUSED(nb30);
                 const uint64_t nb31 = src3->nb[1];
 
                 const int64_t  ne40 = src4->ne[0]; GGML_UNUSED(ne40);
-                const int64_t  ne41 = src4->ne[1]; GGML_UNUSED(ne41);
+                const int64_t  ne41 = src4->ne[1];
                 const int64_t  ne42 = src4->ne[2]; GGML_UNUSED(ne42);
+                const int64_t  ne43 = src4->ne[3]; GGML_UNUSED(ne43);
 
-                const uint64_t nb40 = src4->nb[0];
+                const uint64_t nb40 = src4->nb[0]; GGML_UNUSED(nb40);
                 const uint64_t nb41 = src4->nb[1];
                 const uint64_t nb42 = src4->nb[2];
+                const uint64_t nb43 = src4->nb[3];
 
                 const int64_t  ne50 = src5->ne[0]; GGML_UNUSED(ne50);
                 const int64_t  ne51 = src5->ne[1]; GGML_UNUSED(ne51);
                 const int64_t  ne52 = src5->ne[2]; GGML_UNUSED(ne52);
+                const int64_t  ne53 = src5->ne[2]; GGML_UNUSED(ne53);
 
-                const uint64_t nb50 = src5->nb[0];
+                const uint64_t nb50 = src5->nb[0]; GGML_UNUSED(nb50);
                 const uint64_t nb51 = src5->nb[1];
                 const uint64_t nb52 = src5->nb[2];
+                const uint64_t nb53 = src5->nb[3];
+
+                const int64_t  ne60 = src6->ne[0]; GGML_UNUSED(ne60);
+
+                const uint64_t nb60 = src6->nb[0]; GGML_UNUSED(nb60);
 
                 const int64_t d_state      = ne00;
                 const int64_t d_inner      = ne01;
-                const int64_t n_seq_tokens = ne11;
-                const int64_t n_seqs       = ne02;
+                const int64_t n_head       = ne02;
+                const int64_t n_group      = ne41;
+                const int64_t n_seq_tokens = ne12;
+                const int64_t n_seqs       = ne13;
+
+                id<MTLComputePipelineState> pipeline = nil;
 
-                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32].pipeline;
+                if (ne30 == 1) {
+                    // Mamba-2
+                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32_GROUP].pipeline;
+                } else {
+                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32].pipeline;
+                }
 
                 ggml_metal_kargs_ssm_scan args = {
                     /*.d_state =*/ d_state,
                     /*.d_inner =*/ d_inner,
+                    /*.n_head =*/ n_head,
+                    /*.n_group =*/ n_group,
                     /*.n_seq_tokens =*/ n_seq_tokens,
                     /*.n_seqs =*/ n_seqs,
-                    /*.nb00 =*/ nb00,
                     /*.nb01 =*/ nb01,
                     /*.nb02 =*/ nb02,
-                    /*.nb10 =*/ nb10,
+                    /*.nb03 =*/ nb03,
                     /*.nb11 =*/ nb11,
                     /*.nb12 =*/ nb12,
                     /*.nb13 =*/ nb13,
-                    /*.nb20 =*/ nb20,
                     /*.nb21 =*/ nb21,
                     /*.nb22 =*/ nb22,
-                    /*.nb30 =*/ nb30,
                     /*.nb31 =*/ nb31,
-                    /*.nb40 =*/ nb40,
                     /*.nb41 =*/ nb41,
                     /*.nb42 =*/ nb42,
-                    /*.nb50 =*/ nb50,
+                    /*.nb43 =*/ nb43,
                     /*.nb51 =*/ nb51,
                     /*.nb52 =*/ nb52,
+                    /*.nb53 =*/ nb53,
                 };
 
                 [encoder setComputePipelineState:pipeline];
@@ -2660,10 +2682,18 @@ static bool ggml_metal_encode_node(
                 [encoder setBuffer:id_src3 offset:offs_src3 atIndex:3];
                 [encoder setBuffer:id_src4 offset:offs_src4 atIndex:4];
                 [encoder setBuffer:id_src5 offset:offs_src5 atIndex:5];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:6];
-                [encoder setBytes:&args    length:sizeof(args) atIndex:7];
+                [encoder setBuffer:id_src6 offset:offs_src6 atIndex:6];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:7];
+                [encoder setBytes:&args    length:sizeof(args) atIndex:8];
+                // NOTE: max index is 31
 
-                [encoder dispatchThreadgroups:MTLSizeMake(d_inner, n_seqs, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                if (ne30 == 1) {
+                    // Mamba-2
+                    [encoder dispatchThreadgroups:MTLSizeMake(d_inner, n_head, n_seqs) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } else {
+                    GGML_ASSERT(d_inner == 1);
+                    [encoder dispatchThreadgroups:MTLSizeMake(n_head, n_seqs, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                }
             } break;
         case GGML_OP_RWKV_WKV6:
             {