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

* origin/compilade/mamba2: (24 commits)
kv-cache : allow context shift for recurrent models
convert : avoid AutoConfig for Mamba and Mamba2 hparams
kv-cache : remove const_cast when setting inputs for s_copy
metal : single-user mamba2 inference works
metal : add missing args for nb references in ssm_scan_f32_group
metal : fix confusion between ; and ,
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
...

Author: gabe-l-hart

convert_hf_to_gguf.py

@@ -4492,6 +4492,14 @@ def set_gguf_parameters(self):
 class MambaModel(TextModel):
     model_arch = gguf.MODEL_ARCH.MAMBA
 
+    def __init__(self, dir_model: Path, *args, **kwargs):
+        # Avoid using AutoConfig for hparams
+        hparams = kwargs.pop("hparams", None)
+        if hparams is None:
+            with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+                hparams = json.load(f)
+        super().__init__(dir_model, *args, hparams=hparams, **kwargs)
+
     def set_vocab(self):
         vocab_size = self.hparams["vocab_size"]
         # Round vocab size to next multiple of 8
@@ -4566,6 +4574,100 @@ 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 __init__(self, dir_model: Path, *args, **kwargs):
+        # Avoid using AutoConfig for hparams
+        # It wrongly assumes all Mamba2 models are Mamba-Codestral-7B-v0.1
+        hparams = kwargs.pop("hparams", None)
+        if hparams is None:
+            with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+                hparams = json.load(f)
+        super().__init__(dir_model, *args, hparams=hparams, **kwargs)
+
+    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
@@ -6332,12 +6434,20 @@ def get_model_architecture(hparams: dict[str, Any], model_type: ModelType) -> st
     # maybe we should fallback to text model's arch in that case, since not many models have both
     text_config = hparams.get("text_config", {})
     vision_config = hparams.get("vision_config", {})
-    arch = hparams["architectures"][0]
+    arch = None
+    if (arches := hparams.get("architectures")) is not None and len(arches) > 0:
+        arch = arches[0]
+    elif "ssm_cfg" in hparams:
+        # For non-hf Mamba and Mamba2 models
+        arch = hparams["ssm_cfg"].get("layer", "Mamba") + "ForCausalLM"
+
     # if "architectures" is found in the sub-config, use that instead
     if model_type == ModelType.TEXT and text_config.get("architectures") is not None:
         arch = text_config["architectures"][0]
     elif model_type == ModelType.MMPROJ and vision_config.get("architectures") is not None:
         arch = vision_config["architectures"][0]
+    if arch is None:
+        raise ValueError("Failed to detect model architecture")
     return arch
 
 

ggml/include/ggml.h

@@ -1878,7 +1878,8 @@ 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-cpu/ops.cpp

@@ -7596,74 +7596,151 @@ void ggml_compute_forward_ssm_conv(
 static void ggml_compute_forward_ssm_scan_f32(
         const ggml_compute_params * params,
         ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0]; // s
-    const ggml_tensor * src1 = dst->src[1]; // x
-    const ggml_tensor * src2 = dst->src[2]; // dt
-    const ggml_tensor * src3 = dst->src[3]; // A
-    const ggml_tensor * src4 = dst->src[4]; // B
-    const ggml_tensor * src5 = dst->src[5]; // C
+    const ggml_tensor * src0 = dst->src[0]; // s  {d_state, dim, n_head, n_seqs+}
+    const ggml_tensor * src1 = dst->src[1]; // x  {dim, n_head, n_seq_tokens, n_seqs}
+    const ggml_tensor * src2 = dst->src[2]; // dt {n_head, n_seq_tokens, n_seqs}
+    const ggml_tensor * src3 = dst->src[3]; // A  {d_state, n_head} or {1, n_head}
+    const ggml_tensor * src4 = dst->src[4]; // B  {d_state, n_group, n_seq_tokens, n_seqs}
+    const ggml_tensor * src5 = dst->src[5]; // C  {d_state, n_group, n_seq_tokens, n_seqs}
+    const ggml_tensor * src6 = dst->src[6]; // ids {n_seqs}
 
     const int ith = params->ith;
     const int nth = params->nth;
 
-    const int64_t nc  = src0->ne[0]; // d_state
-    const int64_t nr  = src0->ne[1]; // d_inner
-    const int64_t n_t = src1->ne[1]; // number of tokens per sequence
-    const int64_t n_s = src0->ne[2]; // number of sequences in the batch
+    const int64_t nc = src0->ne[0]; // d_state
+    const int64_t nr = src0->ne[1]; // dim
+    const int64_t nh = src1->ne[1]; // n_head
+    const int64_t ng = src4->ne[1];
+    const int64_t nt = src1->ne[2]; // number of tokens per sequence
+    const int64_t ns = src1->ne[3]; // number of sequences in the batch
 
-    GGML_ASSERT(ggml_nelements(src1) + ggml_nelements(src0) == ggml_nelements(dst));
+    // can't use ggml_nbytes because src1 is not necessarily contiguous
+    const int64_t s_off = ggml_nelements(src1) * ggml_element_size(src1);
+
+    GGML_ASSERT(ggml_nelements(src1) + nc*nr*nh*ns == ggml_nelements(dst));
     GGML_ASSERT(src0->nb[0] == sizeof(float));
     GGML_ASSERT(src1->nb[0] == sizeof(float));
     GGML_ASSERT(src2->nb[0] == sizeof(float));
     GGML_ASSERT(src3->nb[0] == sizeof(float));
     GGML_ASSERT(src4->nb[0] == sizeof(float));
     GGML_ASSERT(src5->nb[0] == sizeof(float));
-    // required for the dot product between s and C
-    GGML_ASSERT(src0->nb[1] == src0->ne[0]*sizeof(float));
-    // required for per-sequence offsets for states
-    GGML_ASSERT(src0->nb[2] == src0->ne[0]*src0->ne[1]*sizeof(float));
-    // required to get correct offset for state destination (i.e. src1->nb[3])
-    GGML_ASSERT(src1->nb[3] == src1->ne[0]*src1->ne[1]*src1->ne[2]*sizeof(float));
+    GGML_ASSERT(src6->nb[0] == sizeof(int32_t));
+    // allows optimizing the modulo since n_group should be a power of 2
+    GGML_ASSERT((ng & -ng) == ng);
+
+    // heads per thread
+    const int dh = (nh + nth - 1)/nth;
+
+    // head range for this thread
+    const int ih0 = dh*ith;
+    const int ih1 = MIN(ih0 + dh, nh);
+
+    const int32_t * ids = (const int32_t *) src6->data;
+
+    for (int i3 = 0; i3 < ns; ++i3) {
+        const float * s0 = (const float *) ((const char *) src0->data + ids[i3]*(src0->nb[3])); // {d_state, dim, nh, ns}
+              float * s  = (      float *) ((      char *) dst->data  + i3*(src0->nb[3]) + s_off); // {d_state, dim, nh, ns}
+
+        for (int i2 = 0; i2 < nt; ++i2) {
+            const float * x  = (const float *) ((const char *) src1->data + i2*(src1->nb[2]) + i3*(src1->nb[3])); // {dim, nh, nt, ns}
+            const float * dt = (const float *) ((const char *) src2->data + i2*(src2->nb[1]) + i3*(src2->nb[2])); // {nh, nt, ns}
+            const float * A  = (const float *) ((const char *) src3->data); // {d_state, nh} or {1, nh}
+            const float * B  = (const float *) ((const char *) src4->data + i2*(src4->nb[2]) + i3*(src4->nb[3])); // {d_state, ng, nt, ns}
+            const float * C  = (const float *) ((const char *) src5->data + i2*(src5->nb[2]) + i3*(src5->nb[3])); // {d_state, ng, nt, ns}
+                  float * y  = (      float *) ((      char *) dst->data + i2*(nh*nr*sizeof(float)) + i3*(nt*nh*nr*sizeof(float))); // {dim, nh, nt, ns}
+
+            if (src3->ne[0] == 1) {
+                // Mamba-2 has a scalar decay factor per head; dA can be outside the state-wise loop
+
+                // n_head
+                for (int h = ih0; h < ih1; ++h) {
+                    // ref: https://github.com/state-spaces/mamba/blob/62db608da60f6fc790b8ed9f4b3225e95ca15fde/mamba_ssm/ops/triton/softplus.py#L16
+                    const float dt_soft_plus = dt[h] <= 20.0f ? log1pf(expf(dt[h])) : dt[h];
+                    const float dA = expf(dt_soft_plus * A[h]);
+
+                    // dim
+                    for (int i1 = 0; i1 < nr; ++i1) {
+                        const int ii = i1 + h*nr;
+                        const float x_dt = x[ii] * dt_soft_plus;
+                        float sumf = 0.0f;
+#if defined(GGML_SIMD)
+                        const int np = (nc & ~(GGML_F32_STEP - 1));
 
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
+                        GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
 
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
-    const int ir  = ir1 - ir0;
+                        GGML_F32_VEC adA = GGML_F32_VEC_SET1(dA);
+                        GGML_F32_VEC axdt = GGML_F32_VEC_SET1(x_dt);
 
-    for (int i3 = 0; i3 < n_s; ++i3) {
-        for (int i2 = 0; i2 < n_t; ++i2) {
-            const float * s0 = (const float *) ((const char *) src0->data + ir0*(src0->nb[1]) + i3*(src0->nb[2])); // {d_state, d_inner, n_s}
-            const float * x  = (const float *) ((const char *) src1->data + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s}
-            const float * dt = (const float *) ((const char *) src2->data + ir0*(src2->nb[0]) + i2*(src2->nb[1]) + i3*(src2->nb[2])); // {d_inner, n_t, n_s}
-            const float * A  = (const float *) ((const char *) src3->data + ir0*(src3->nb[1])); // {d_state, d_inner}
-            const float * B  = (const float *) ((const char *) src4->data +  i2*(src4->nb[1]) + i3*(src4->nb[2])); // {d_state, n_t, n_s}
-            const float * C  = (const float *) ((const char *) src5->data +  i2*(src5->nb[1]) + i3*(src5->nb[2])); // {d_state, n_t, n_s}
-                  float * y  = (      float *) ((      char *) dst->data  + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s}
-                  float * s  = (      float *) ((      char *) dst->data  + ir0*(src0->nb[1]) + i3*(src0->nb[2]) +     src1->nb[3]);  // {d_state, d_inner, n_s}
-
-            // use the output as the source for the next token-wise iterations
-            if (i2 > 0) { s0 = s; }
+                        GGML_F32_VEC ax[GGML_F32_ARR];
+                        GGML_F32_VEC ay[GGML_F32_ARR];
+                        GGML_F32_VEC az[GGML_F32_ARR];
 
-            // d_inner
-            for (int i1 = 0; i1 < ir; ++i1) {
-                // ref: https://github.com/state-spaces/mamba/blob/34076d664838588a3c97727b263478ab9f621a07/mamba_ssm/ops/triton/selective_state_update.py#L78
-                float dt_soft_plus = dt[i1] <= 20.0f ? log1pf(expf(dt[i1])) : dt[i1];
-                float x_dt = x[i1] * dt_soft_plus;
-                float sumf = 0.0f;
-                // d_state
-                for (int i0 = 0; i0 < nc; ++i0) {
-                    int i = i0 + i1*nc;
-                    // state = prev_state * dA + dB * x
-                    float state = (s0[i] * expf(dt_soft_plus * A[i])) + (B[i0] * x_dt);
-                    // y = rowwise_dotprod(state, C)
-                    sumf += state * C[i0];
-                    s[i] = state;
+                        for (int i = 0; i < np; i += GGML_F32_STEP) {
+                            for (int j = 0; j < GGML_F32_ARR; j++) {
+                                ax[j] = GGML_F32_VEC_LOAD(s0 + i + j*GGML_F32_EPR + ii*nc);
+                                ay[j] = GGML_F32_VEC_LOAD(B + i + j*GGML_F32_EPR + (h & (ng - 1))*nc);
+                                az[j] = GGML_F32_VEC_LOAD(C + i + j*GGML_F32_EPR + (h & (ng - 1))*nc);
+
+                                ax[j] = GGML_F32_VEC_MUL(ax[j], adA);
+                                ay[j] = GGML_F32_VEC_MUL(ay[j], axdt);
+
+                                ax[j] = GGML_F32_VEC_ADD(ax[j], ay[j]);
+
+                                sum[j] = GGML_F32_VEC_FMA(sum[j], ax[j], az[j]);
+
+                                GGML_F32_VEC_STORE(s + i + j*GGML_F32_EPR + ii*nc, ax[j]);
+                            }
+                        }
+
+                        // reduce sum0..sum3 to sum0
+                        GGML_F32_VEC_REDUCE(sumf, sum);
+#else
+                        const int np = 0;
+#endif
+                        // d_state
+                        for (int i0 = np; i0 < nc; ++i0) {
+                            const int i = i0 + ii*nc;
+                            const int ig = i0 + (h & (ng - 1))*nc;
+                            // state = prev_state * dA + dB * x
+                            const float state = (s0[i] * dA) + (B[ig] * x_dt);
+                            // y = rowwise_dotprod(state, C)
+                            sumf += state * C[ig];
+                            s[i] = state;
+                        }
+                        y[ii] = sumf;
+                    }
+                }
+            } else {
+                // Mamba-1 has an element-wise decay factor for the states
+
+                // n_head
+                for (int h = ih0; h < ih1; ++h) {
+                    // ref: https://github.com/state-spaces/mamba/blob/62db608da60f6fc790b8ed9f4b3225e95ca15fde/mamba_ssm/ops/triton/softplus.py#L16
+                    const float dt_soft_plus = dt[h] <= 20.0f ? log1pf(expf(dt[h])) : dt[h];
+
+                    // dim
+                    for (int i1 = 0; i1 < nr; ++i1) {
+                        const int ii = i1 + h*nr;
+                        const float x_dt = x[ii] * dt_soft_plus;
+                        float sumf = 0.0f;
+                        // NOTE: can't really use GGML_SIMD here because d_state is usually 16
+                        //       and also because expf is used within the loop.
+                        // d_state
+                        for (int i0 = 0; i0 < nc; ++i0) {
+                            const int i = i0 + ii*nc;
+                            const int ig = i0 + (h & (ng - 1))*nc;
+                            // state = prev_state * dA + dB * x
+                            const float state = (s0[i] * expf(dt_soft_plus * A[i0 + h*nc])) + (B[ig] * x_dt);
+                            // y = rowwise_dotprod(state, C)
+                            sumf += state * C[ig];
+                            s[i] = state;
+                        }
+                        y[ii] = sumf;
+                    }
                 }
-                y[i1] = sumf;
             }
+            // use the output as the source when it's not the first token-wise iteration
+            s0 = s;
         }
     }
 }

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

@@ -488,26 +488,25 @@ 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 {