Merge branch 'ml-explore:main' into adding-support-for-mamba2

llms/mlx_lm/models/deepseek_v2.py

@@ -282,12 +282,12 @@ class MoEGate(nn.Module):
         if self.topk_method == "group_limited_greedy":
             bsz, seq_len = x.shape[:2]
             scores = scores.reshape(bsz, seq_len, self.n_group, -1)
-            group_scores = scores.max(axis=-1)
+            group_scores = scores.max(axis=-1, keepdims=True)
             k = self.n_group - self.topk_group
-            group_idx = mx.argpartition(group_scores, kth=k - 1, axis=-1)[..., :k]
-            batch_idx = mx.expand_dims(mx.arange(bsz), (1, 2))
-            seq_idx = mx.expand_dims(mx.arange(seq_len), (0, 2))
-            scores[batch_idx, seq_idx, group_idx] = 0.0
+            group_idx = mx.argpartition(group_scores, kth=k - 1, axis=-2)[..., :k, :]
+            scores = mx.put_along_axis(
+                scores, group_idx, mx.array(0.0, scores.dtype), axis=-2
+            )
             scores = scores.reshape(bsz, seq_len, -1)
 
         k = self.top_k
@@ -364,8 +364,32 @@ class DeepseekV2Model(nn.Module):
             DeepseekV2DecoderLayer(config, idx)
             for idx in range(config.num_hidden_layers)
         ]
+        self.start_idx = 0
+        self.end_idx = len(self.layers)
+        self.num_layers = self.end_idx
+
         self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
 
+        self.pipeline_rank = 0
+        self.pipeline_size = 1
+
+    def pipeline(self, group):
+        # Split layers in reverse so rank=0 gets the last layers and
+        # rank=pipeline_size-1 gets the first
+        self.pipeline_rank = group.rank()
+        self.pipeline_size = group.size()
+        layers_per_rank = len(self.layers) // self.pipeline_size
+        extra = len(self.layers) - layers_per_rank * self.pipeline_size
+        if self.pipeline_rank < extra:
+            layers_per_rank += 1
+
+        self.start_idx = (self.pipeline_size - self.pipeline_rank - 1) * layers_per_rank
+        self.end_idx = self.start_idx + layers_per_rank
+        self.num_layers = layers_per_rank
+        self.layers = self.layers[: self.end_idx]
+        self.layers[: self.start_idx] = [None] * self.start_idx
+        self.num_layers = len(self.layers) - self.start_idx
+
     def __call__(
         self,
         x: mx.array,
@@ -374,14 +398,31 @@ class DeepseekV2Model(nn.Module):
     ) -> mx.array:
         h = self.embed_tokens(x)
 
+        pipeline_rank = self.pipeline_rank
+        pipeline_size = self.pipeline_size
+        # Hack to avoid time-outs during prompt-processing
+        dist_stream = mx.cpu if h.shape[1] > 1 else mx.gpu
         if mask is None:
             mask = create_attention_mask(h, cache)
 
         if cache is None:
-            cache = [None] * len(self.layers)
+            cache = [None] * self.num_layers
 
-        for layer, c in zip(self.layers, cache):
-            h = layer(h, mask, c)
+        # Receive from the previous process in the pipeline
+        if pipeline_rank < pipeline_size - 1:
+            h = mx.distributed.recv_like(h, (pipeline_rank + 1), stream=dist_stream)
+
+        for i in range(self.num_layers):
+            h = self.layers[self.start_idx + i](h, mask, cache[i])
+
+        # Send to the next process in the pipeline
+        if pipeline_rank != 0:
+            h = mx.distributed.send(
+                h, (pipeline_rank - 1) % pipeline_size, stream=dist_stream
+            )
+
+        # Broadcast h while keeping it in the graph
+        h = mx.distributed.all_gather(h, stream=dist_stream)[: h.shape[0]]
 
         return self.norm(h)
 
@@ -418,4 +459,4 @@ class Model(nn.Module):
 
     @property
     def layers(self):
-        return self.model.layers
+        return self.model.layers[self.model.start_idx : self.model.end_idx]

llms/mlx_lm/models/deepseek_v3.py

@@ -271,6 +271,38 @@ class DeepseekV3MLP(nn.Module):
         return down_proj
 
 
+@mx.compile
+def group_expert_select(
+    gates,
+    e_score_correction_bias,
+    top_k,
+    n_group,
+    topk_group,
+    routed_scaling_factor,
+    norm_topk_prob,
+):
+
+    k = top_k
+    scores = mx.sigmoid(gates.astype(mx.float32))
+    scores = scores + e_score_correction_bias
+    scores = mx.unflatten(scores, axis=-1, shape=(n_group, -1))
+    group_scores = mx.topk(scores, 2, axis=-1).sum(axis=-1, keepdims=True)
+    k = n_group - topk_group
+    group_idx = mx.argpartition(group_scores, kth=k - 1, axis=-2)[..., :k, :]
+    scores = mx.put_along_axis(scores, group_idx, mx.array(0.0), axis=-2)
+    scores = mx.flatten(scores, -2, -1)
+
+    k = top_k
+    inds = mx.argpartition(-scores, kth=k - 1, axis=-1)[..., :k]
+    scores = mx.take_along_axis(scores, inds, axis=-1)
+    if top_k > 1 and norm_topk_prob:
+        denominator = scores.sum(axis=-1, keepdims=True) + 1e-20
+        scores = scores / denominator
+    scores = scores * routed_scaling_factor
+
+    return inds, scores
+
+
 class MoEGate(nn.Module):
     def __init__(self, config: ModelArgs):
         super().__init__()
@@ -279,38 +311,22 @@ class MoEGate(nn.Module):
         self.norm_topk_prob = config.norm_topk_prob
         self.n_routed_experts = config.n_routed_experts
         self.routed_scaling_factor = config.routed_scaling_factor
-        self.topk_method = config.topk_method
         self.n_group = config.n_group
         self.topk_group = config.topk_group
         self.weight = mx.zeros((self.n_routed_experts, config.hidden_size))
         self.e_score_correction_bias = mx.zeros((self.n_routed_experts,))
+        assert config.topk_method == "noaux_tc", "Unsupported topk method."
 
     def __call__(self, x):
-        gates = x @ self.weight.T
-
-        scores = mx.sigmoid(gates.astype(mx.float32))
-
-        assert self.topk_method == "noaux_tc", "Unsupported topk method."
-        bsz, seq_len = x.shape[:2]
-        scores = scores + self.e_score_correction_bias
-        scores = scores.reshape(bsz, seq_len, self.n_group, -1)
-        group_scores = mx.topk(scores, 2, axis=-1).sum(axis=-1)
-        k = self.n_group - self.topk_group
-        group_idx = mx.argpartition(group_scores, kth=k - 1, axis=-1)[..., :k]
-        batch_idx = mx.expand_dims(mx.arange(bsz), (1, 2))
-        seq_idx = mx.expand_dims(mx.arange(seq_len), (0, 2))
-        scores[batch_idx, seq_idx, group_idx] = 0.0
-        scores = scores.reshape(bsz, seq_len, -1)
-
-        k = self.top_k
-        inds = mx.argpartition(-scores, kth=k - 1, axis=-1)[..., :k]
-        scores = mx.take_along_axis(scores, inds, axis=-1)
-        if self.top_k > 1 and self.norm_topk_prob:
-            denominator = scores.sum(axis=-1, keepdims=True) + 1e-20
-            scores = scores / denominator
-        scores = scores * self.routed_scaling_factor
-
-        return inds, scores
+        return group_expert_select(
+            x @ self.weight.T,
+            self.e_score_correction_bias,
+            self.top_k,
+            self.n_group,
+            self.topk_group,
+            self.routed_scaling_factor,
+            self.norm_topk_prob,
+        )
 
 
 class DeepseekV3MoE(nn.Module):
@@ -381,6 +397,10 @@ class DeepseekV3Model(nn.Module):
             DeepseekV3DecoderLayer(config, idx)
             for idx in range(config.num_hidden_layers)
         ]
+        self.start_idx = 0
+        self.end_idx = len(self.layers)
+        self.num_layers = self.end_idx
+
         self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
         self.pipeline_rank = 0
         self.pipeline_size = 1
@@ -390,11 +410,15 @@ class DeepseekV3Model(nn.Module):
         # rank=pipeline_size-1 gets the first
         self.pipeline_rank = group.rank()
         self.pipeline_size = group.size()
-        layers_per_rank = (
-            len(self.layers) + self.pipeline_size - 1
-        ) // self.pipeline_size
-        start = (self.pipeline_size - self.pipeline_rank - 1) * layers_per_rank
-        self.layers = self.layers[start : start + layers_per_rank]
+        layers_per_rank = len(self.layers) // self.pipeline_size
+        extra = len(self.layers) - layers_per_rank * self.pipeline_size
+        if self.pipeline_rank < extra:
+            layers_per_rank += 1
+        self.start_idx = (self.pipeline_size - self.pipeline_rank - 1) * layers_per_rank
+        self.end_idx = self.start_idx + layers_per_rank
+        self.layers = self.layers[: self.end_idx]
+        self.layers[: self.start_idx] = [None] * self.start_idx
+        self.num_layers = len(self.layers) - self.start_idx
 
     def __call__(
         self,
@@ -412,15 +436,15 @@ class DeepseekV3Model(nn.Module):
             mask = create_attention_mask(h, cache)
 
         if cache is None:
-            cache = [None] * len(self.layers)
+            cache = [None] * self.num_layers
 
         # Receive from the previous process in the pipeline
 
         if pipeline_rank < pipeline_size - 1:
             h = mx.distributed.recv_like(h, (pipeline_rank + 1), stream=dist_stream)
 
-        for layer, c in zip(self.layers, cache):
-            h = layer(h, mask, c)
+        for i in range(self.num_layers):
+            h = self.layers[self.start_idx + i](h, mask, cache[i])
 
         # Send to the next process in the pipeline
         if pipeline_rank != 0:
@@ -468,4 +492,4 @@ class Model(nn.Module):
 
     @property
     def layers(self):
-        return self.model.layers
+        return self.model.layers[self.model.start_idx : self.model.end_idx]

llms/mlx_lm/models/granite.py

@@ -0,0 +1,195 @@
+# Copyright © 2023-2024 Apple Inc.
+
+from dataclasses import dataclass
+from typing import Any, Dict, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .rope_utils import initialize_rope
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    rms_norm_eps: float
+    vocab_size: int
+    logits_scaling: float
+    attention_multiplier: float
+    embedding_multiplier: float
+    residual_multiplier: float
+    max_position_embeddings: int
+    num_key_value_heads: int
+    attention_bias: bool
+    mlp_bias: bool
+    rope_theta: float
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+    tie_word_embeddings: bool = True
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+
+        self.head_dim = head_dim = args.hidden_size // n_heads
+
+        self.scale = args.attention_multiplier
+        attention_bias = args.attention_bias
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=attention_bias)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=attention_bias)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=attention_bias)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=attention_bias)
+
+        self.rope = initialize_rope(
+            self.head_dim,
+            args.rope_theta,
+            False,
+            args.rope_scaling,
+            args.max_position_embeddings,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        # Prepare the queries, keys and values for the attention computation
+        queries = queries.reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class MLP(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_size
+        hidden_dim = args.intermediate_size
+        if hasattr(args, "mlp_bias"):
+            mlp_bias = args.mlp_bias
+        else:
+            mlp_bias = False
+
+        self.gate_proj = nn.Linear(dim, hidden_dim, bias=mlp_bias)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=mlp_bias)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=mlp_bias)
+
+    def __call__(self, x) -> mx.array:
+        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.num_attention_heads = args.num_attention_heads
+        self.hidden_size = args.hidden_size
+        self.self_attn = Attention(args)
+        self.mlp = MLP(args)
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+        self.residual_multiplier = args.residual_multiplier
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r * self.residual_multiplier
+        r = self.mlp(self.post_attention_layernorm(h))
+        out = h + r * self.residual_multiplier
+        return out
+
+
+class GraniteModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+        self.num_hidden_layers = args.num_hidden_layers
+        assert self.vocab_size > 0
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            TransformerBlock(args=args) for _ in range(args.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.embedding_multiplier = args.embedding_multiplier
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        h = self.embed_tokens(inputs) * self.embedding_multiplier
+
+        if mask is None:
+            mask = create_attention_mask(h, cache)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, cache=c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = GraniteModel(args)
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+        self.logits_scaling = args.logits_scaling
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        out = self.model(inputs, mask, cache)
+        if self.args.tie_word_embeddings:
+            out = self.model.embed_tokens.as_linear(out)
+        else:
+            out = self.lm_head(out)
+        return out / self.logits_scaling
+
+    @property
+    def layers(self):
+        return self.model.layers

llms/mlx_lm/models/helium.py

@@ -1,3 +1,5 @@
+# Copyright © 2025 Apple Inc.
+
 from dataclasses import dataclass
 from typing import Any, Optional, Tuple
 

llms/mlx_lm/models/hunyuan.py

@@ -76,7 +76,6 @@ class Attention(nn.Module):
 
         head_dim = args.hidden_size // n_heads
         self.scale = head_dim**-0.5
-
         self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=args.attention_bias)
         if kv_proj:
             self.k_proj = nn.Linear(
@@ -107,7 +106,6 @@ class Attention(nn.Module):
         B, L, D = x.shape
 
         queries = self.q_proj(x)
-
         if kv_states is None:
             keys, values = self.k_proj(x), self.v_proj(x)
             kv_states = keys, values
@@ -198,7 +196,10 @@ class DecoderLayer(nn.Module):
         super().__init__()
         self.hidden_size = args.hidden_size
         self.self_attn = Attention(kv_proj, args)
-        self.mlp = MoeBlock(args)
+        if args.num_experts == 1:
+            self.mlp = MLP(args.hidden_size, args.intermediate_size)
+        else:
+            self.mlp = MoeBlock(args)
 
         self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
         self.post_attention_layernorm = nn.RMSNorm(
@@ -231,7 +232,10 @@ class HunYuanModel(nn.Module):
         assert self.vocab_size > 0
         self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
         self.layers = [
-            DecoderLayer(args=args, kv_proj=(i % args.cla_share_factor) == 0)
+            DecoderLayer(
+                args=args,
+                kv_proj=(not args.use_cla) or (i % args.cla_share_factor) == 0,
+            )
             for i in range(args.num_hidden_layers)
         ]
         self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
@@ -251,7 +255,7 @@ class HunYuanModel(nn.Module):
             cache = [None] * len(self.layers)
 
         for i, (layer, c) in enumerate(zip(self.layers, cache)):
-            if i % self.args.cla_share_factor == 0:
+            if (not self.args.use_cla) or i % self.args.cla_share_factor == 0:
                 shared_kv_states = None
             h, shared_kv_states = layer(h, mask, c, shared_kv_states)
 
@@ -275,6 +279,29 @@ class Model(nn.Module):
         return self.model.embed_tokens.as_linear(out)
 
     def sanitize(self, weights):
+
+        if "model.layers.0.mlp.gate_and_up_proj.weight" in weights:
+            new_weights = {}
+            D = self.args.hidden_size
+            n_kv_heads = self.args.num_key_value_heads
+            n_kv_groups = self.args.num_attention_heads // n_kv_heads
+            head_dim = D // self.args.num_attention_heads
+            for k, v in weights.items():
+                if "qkv_proj" in k:
+                    v = v.reshape(n_kv_heads, n_kv_groups + 2, head_dim, -1)
+                    splits = v.split([n_kv_groups, n_kv_groups + 1], axis=1)
+                    for k_up, v_new in zip(["q_proj", "k_proj", "v_proj"], splits):
+                        k_new = k.replace("qkv_proj", k_up)
+                        new_weights[k_new] = mx.flatten(v_new, 0, 2)
+                elif "gate_and_up_proj" in k:
+                    splits = v.split(2, axis=0)
+                    for k_up, v_new in zip(["up_proj", "gate_proj"], splits):
+                        k_new = k.replace("gate_and_up_proj", k_up)
+                        new_weights[k_new] = v_new
+                else:
+                    new_weights[k] = v
+            weights = new_weights
+
         if "model.layers.0.mlp.experts.0.up_proj.weight" not in weights:
             return weights
         for l in range(self.args.num_hidden_layers):

llms/mlx_lm/models/mamba.py

@@ -1,4 +1,4 @@
-# Copyright © 2024 Apple Inc.
+# Copyright © 2024-2025 Apple Inc.
 
 import math
 from dataclasses import dataclass
@@ -123,17 +123,16 @@ class MambaBlock(nn.Module):
             self.intermediate_size, self.hidden_size, bias=args.use_bias
         )
 
-    def ssm_step(self, x, state=None):
-        A = -mx.exp(self.A_log)
+    def ssm_step(self, x, A, state=None):
         D = self.D
         deltaBC = self.x_proj(x)
-        delta, B, C = mx.split(
-            deltaBC,
-            indices_or_sections=[
-                self.time_step_rank,
-                self.time_step_rank + self.ssm_state_size,
-            ],
-            axis=-1,
+        delta, B, C = map(
+            self.mixer_norm if self.use_bcdt_rms else lambda x: x,
+            mx.split(
+                deltaBC,
+                [self.time_step_rank, self.time_step_rank + self.ssm_state_size],
+                axis=-1,
+            ),
         )
         if self.use_bcdt_rms:
             delta, B, C = map(self.mixer_norm, (delta, B, C))
@@ -145,25 +144,40 @@ class MambaBlock(nn.Module):
         y = y + D * x
         return y, new_state
 
-    def __call__(self, x, cache):
+    def _process_sequence(self, x, conv_cache, state_cache):
         B, T, D = x.shape
-        if cache is None:
-            cache = [None, None]
+        xz = self.in_proj(x)
+        x, z = xz.split(indices_or_sections=2, axis=-1)
+
+        conv_out, new_conv_cache = self.conv1d(x, conv_cache)
+        x = nn.silu(conv_out)
+
+        A = -mx.exp(self.A_log)
 
         outputs = []
+        current_state = state_cache
+        y = []
         for t in range(T):
-            xt = x[:, t, :]
-            xz = self.in_proj(xt)
-            x_t, z_t = xz.split(indices_or_sections=2, axis=1)
-            conv_out, cache[0] = self.conv1d(mx.expand_dims(x_t, 1), cache[0])
-            x_t = conv_out.squeeze(1)
-            x_t = nn.silu(x_t)
-            y_t, cache[1] = self.ssm_step(x_t, cache[1])
-            z_t = nn.silu(z_t)
-            output_t = y_t * z_t
-            output_t = self.out_proj(output_t)
-            outputs.append(output_t)
-        output = mx.stack(outputs, axis=1)
+            y_t, current_state = self.ssm_step(x[:, t], A, current_state)
+            y.append(y_t)
+        y = mx.stack(y, axis=1)
+        z = self.out_proj(nn.silu(z) * y)
+        return z, (new_conv_cache, current_state)
+
+    def __call__(self, x, cache):
+        if cache is None:
+            conv_cache, state_cache = None, None
+        else:
+            conv_cache, state_cache = cache[0], cache[1]
+
+        output, (new_conv_cache, new_state_cache) = self._process_sequence(
+            x, conv_cache, state_cache
+        )
+
+        if isinstance(cache, MambaCache):
+            cache[0] = new_conv_cache
+            cache[1] = new_state_cache
+
         return output
 
 

llms/mlx_lm/models/minicpm.py

@@ -1,4 +1,4 @@
-# Copyright © 2023-2024 Apple Inc.
+# Copyright © 2023-2025 Apple Inc.
 
 from dataclasses import dataclass
 from typing import Any, Dict, Optional, Tuple, Union