Starting to get there.

server/text_generation_server/models/custom_modeling/mllama.py

@@ -22,7 +22,6 @@ from torch import nn
 import flash_attn_2_cuda
 
 from transformers.activations import ACT2FN
-from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
 import torch.nn.functional as F
 
 from text_generation_server.layers import (
@@ -734,6 +733,7 @@ class MllamaTextCrossAttention(nn.Module):
             cu_seqlen_k,
             max_q,
             max_k,
+            indices,
         ) = cross_attention_states
 
         key_states = self.k_proj(cross_attention_states)
@@ -862,6 +862,8 @@ class FlashLlamaCrossLayer(torch.nn.Module):
             return hidden_states, residual
         if residual is not None:
             hidden_states += residual
+
+        # indices = cross_attention_states[-1]
         if cu_seqlen_prefill is not None:
             out_hidden_states = hidden_states[:]
             hidden_states = hidden_states[:]
@@ -892,115 +894,6 @@ class FlashLlamaCrossLayer(torch.nn.Module):
         return hidden_states, None
 
 
-class MllamaTextSelfAttention(nn.Module):
-    def __init__(self, *, prefix, config, weights, layer_idx):
-        super().__init__()
-        self.config = config
-        self.num_heads = config.num_attention_heads
-        self.dropout = config.dropout
-        self.hidden_size = config.hidden_size
-        self.num_key_value_heads = config.num_key_value_heads
-        self.head_dim = config.hidden_size // self.num_heads
-        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
-
-        self.num_heads = self.num_heads // weights.process_group.size()
-        self.num_key_value_heads = (
-            self.num_key_value_heads // weights.process_group.size()
-        )
-        self.layer_idx = layer_idx
-
-        self.qkv_proj = TensorParallelColumnLinear.load_multi(
-            config,
-            prefixes=[f"{prefix}.q_proj", f"{prefix}.k_proj", f"{prefix}.v_proj"],
-            dim=0,
-            weights=weights,
-            bias=False,
-        )
-        self.o_proj = TensorParallelRowLinear.load(
-            config,
-            prefix=f"{prefix}.o_proj",
-            weights=weights,
-            bias=False,
-        )
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        attention_mask: torch.Tensor,
-        position_embeddings: torch.Tensor,
-        past_key_value=None,
-        cache_position=None,
-        **kwargs,
-    ):
-        bsz, q_len, _ = hidden_states.size()
-
-        qkv = self.qkv_proj(hidden_states)
-        query_states, key_states, value_states = qkv.split(
-            [
-                self.head_dim * self.num_heads,
-                self.head_dim * self.num_key_value_heads,
-                self.head_dim * self.num_key_value_heads,
-            ],
-            dim=2,
-        )
-
-        query_states = query_states.view(
-            bsz, q_len, self.num_heads, self.head_dim
-        ).transpose(1, 2)
-        key_states = key_states.view(
-            bsz, q_len, self.num_key_value_heads, self.head_dim
-        ).transpose(1, 2)
-        value_states = value_states.view(
-            bsz, q_len, self.num_key_value_heads, self.head_dim
-        ).transpose(1, 2)
-
-        cos, sin = position_embeddings
-        query_states, key_states = apply_rotary_pos_emb(
-            query_states, key_states, cos, sin
-        )
-
-        if past_key_value is not None:
-            # sin and cos are specific to RoPE models; cache_position needed for the static cache
-            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
-            key_states, value_states = past_key_value.update(
-                key_states, value_states, self.layer_idx, cache_kwargs
-            )
-
-        key_states = repeat_kv(key_states, self.num_key_value_groups)
-        value_states = repeat_kv(value_states, self.num_key_value_groups)
-
-        causal_mask = attention_mask
-        if attention_mask is not None:
-            causal_mask = causal_mask[:, :, :, : key_states.shape[-2]]
-
-        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
-        # Reference: https://github.com/pytorch/pytorch/issues/112577.
-        if query_states.device.type == "cuda" and causal_mask is not None:
-            query_states = query_states.contiguous()
-            key_states = key_states.contiguous()
-            value_states = value_states.contiguous()
-
-        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
-        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
-        is_causal = True if causal_mask is None and q_len > 1 else False
-
-        attn_output = torch.nn.functional.scaled_dot_product_attention(
-            query_states,
-            key_states,
-            value_states,
-            # TODO
-            # attn_mask=causal_mask,
-            dropout_p=self.dropout if self.training else 0.0,
-            is_causal=is_causal,
-        )
-
-        attn_output = attn_output.transpose(1, 2).contiguous()
-        attn_output = attn_output.view(bsz, q_len, -1)
-
-        attn_output = self.o_proj(attn_output)
-        return attn_output, None, past_key_value
-
-
 # Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->MllamaText
 class MllamaTextRMSNorm(nn.Module):
     def __init__(self, weight, eps):
@@ -1026,144 +919,6 @@ class MllamaTextRMSNorm(nn.Module):
         return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
 
 
-# Copied from transformers.models.llama.modeling_llama.LlamaDecoderLayer with LlamaDecoder->MllamaSelfAttentionDecoder, Llama->MllamaText, LLAMA->MLLAMA_TEXT
-class MllamaSelfAttentionDecoderLayer(nn.Module):
-    def __init__(self, *, prefix, config, weights, layer_idx):
-        super().__init__()
-        self.hidden_size = config.hidden_size
-
-        self.self_attn = MllamaTextSelfAttention(
-            prefix=f"{prefix}.self_attn",
-            config=config,
-            weights=weights,
-            layer_idx=layer_idx,
-        )
-
-        self.mlp = MllamaTextMLP(prefix=f"{prefix}.mlp", config=config, weights=weights)
-        self.input_layernorm = MllamaTextRMSNorm.load(
-            prefix=f"{prefix}.input_layernorm", weights=weights, eps=config.rms_norm_eps
-        )
-        self.post_attention_layernorm = MllamaTextRMSNorm.load(
-            prefix=f"{prefix}.post_attention_layernorm",
-            weights=weights,
-            eps=config.rms_norm_eps,
-        )
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_value=None,
-        cache_position: Optional[torch.LongTensor] = None,
-        position_embeddings: Optional[
-            Tuple[torch.Tensor, torch.Tensor]
-        ] = None,  # will become mandatory in v4.45
-        image_indices: Optional[torch.LongTensor] = None,
-        **kwargs,
-    ) -> Tuple[
-        torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]
-    ]:
-        residual = hidden_states
-
-        hidden_states = self.input_layernorm(hidden_states)
-
-        # Self Attention
-        hidden_states, self_attn_weights, present_key_value = self.self_attn(
-            hidden_states=hidden_states,
-            attention_mask=attention_mask,
-            position_ids=position_ids,
-            past_key_value=past_key_value,
-            cache_position=cache_position,
-            position_embeddings=position_embeddings,
-            **kwargs,
-        )
-        hidden_states = residual + hidden_states
-
-        # Fully Connected
-        residual = hidden_states
-        hidden_states = self.post_attention_layernorm(hidden_states)
-        hidden_states = self.mlp(hidden_states)
-        hidden_states = residual + hidden_states
-
-        return hidden_states
-
-
-class MllamaRotaryEmbedding(nn.Module):
-    def __init__(
-        self,
-        *,
-        config,
-        weights,
-    ):
-        super().__init__()
-        device = weights.device
-        self.rope_type = config.rope_scaling["rope_type"]
-        self.max_seq_len_cached = config.max_position_embeddings
-        self.original_max_seq_len = config.max_position_embeddings
-
-        self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
-        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
-        inv_freq.to(device=device)
-        self.register_buffer("inv_freq", inv_freq, persistent=False)
-        self.original_inv_freq = self.inv_freq
-
-    def _dynamic_frequency_update(self, position_ids, device):
-        """
-        dynamic RoPE layers should recompute `inv_freq` in the following situations:
-        1 - growing beyond the cached sequence length (allow scaling)
-        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
-        """
-        seq_len = torch.max(position_ids) + 1
-        if seq_len > self.max_seq_len_cached:  # growth
-            inv_freq, self.attention_scaling = self.rope_init_fn(
-                self.config, device, seq_len=seq_len, **self.rope_kwargs
-            )
-            self.register_buffer(
-                "inv_freq", inv_freq, persistent=False
-            )  # TODO joao: may break with compilation
-            self.max_seq_len_cached = seq_len
-
-        if (
-            seq_len < self.original_max_seq_len
-            and self.max_seq_len_cached > self.original_max_seq_len
-        ):  # reset
-            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
-            self.max_seq_len_cached = self.original_max_seq_len
-
-    @torch.no_grad()
-    def forward(self, x, position_ids):
-        if "dynamic" in self.rope_type:
-            self._dynamic_frequency_update(position_ids, device=x.device)
-
-        # Core RoPE block
-        inv_freq_expanded = (
-            self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
-        )
-        position_ids_expanded = position_ids[:, None, :].float()
-        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
-        device_type = x.device.type
-        device_type = (
-            device_type
-            if isinstance(device_type, str) and device_type != "mps"
-            else "cpu"
-        )
-
-        freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(
-            1, 2
-        )
-        emb = torch.cat((freqs, freqs), dim=-1)
-        cos = emb.cos()
-        sin = emb.sin()
-
-        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
-        cos = cos * self.attention_scaling
-        sin = sin * self.attention_scaling
-
-        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
-
-
 class MllamaForConditionalGeneration(nn.Module):
     def __init__(self, prefix, config, weights):
         super().__init__()
@@ -1192,14 +947,15 @@ class MllamaForConditionalGeneration(nn.Module):
                 "`aspect_ratio_ids` must be provided if `pixel_values` is provided"
             )
         # logger.info(f"PIxel values {pixel_values.shape}")
+        batch_size = pixel_values.shape[0]
         vision_states = self.vision_model(
             pixel_values, aspect_ratio_ids, aspect_ratio_mask
         )
         cross_attention_states = self.multi_modal_projector(vision_states).reshape(
             -1, vision_states.shape[-2], self.hidden_size
         )
-        n, m, h = cross_attention_states.shape
+        _, _, h = cross_attention_states.shape
-        cross_attention_states = cross_attention_states.view(1, n * m, h)
+        cross_attention_states = cross_attention_states.view(batch_size, -1, h)
         # logger.info(f"cross {cross_attention_states.shape}")
         return cross_attention_states
 
@@ -1219,7 +975,7 @@ class MllamaForConditionalGeneration(nn.Module):
         cross_attention_states: Optional[torch.Tensor],
         image_indices,
     ):
-        # if cross_attention_mask is not None:
+        # if cross_att_sention_mask is not None:
         #     cross_attention_mask, full_text_row_masked_out_mask = (
         #         _prepare_cross_attention_mask(
         #             cross_attention_mask,
@@ -1237,24 +993,59 @@ class MllamaForConditionalGeneration(nn.Module):
         #     ]
 
         if cross_attention_states is not None:
-            seqlen_q = input_ids.shape[0]
+            seqlen_q = len(image_indices)
+            n_images = cross_attention_states.shape[0]
             seqlen_k = cross_attention_states.shape[1]
+            device = cross_attention_states.device
+            if cu_seqlen_prefill is not None:
+                # raise RuntimeError("TODO")
+                offset = 0
+                cu_q = []
+                indices = []
+                for index in image_indices:
+                    cu_q.append(offset)
+                    length = seqlen.input_lengths[index]
+                    input_ids_offset = seqlen.cu_seqlen_q[index]
+                    indices.extend(range(input_ids_offset, input_ids_offset + length))
+                    offset += length
+                cu_q.append(offset)
+                cu_seqlen_q = torch.Tensor(cu_q).to(device=device, dtype=torch.int32)
+
+                cu_seqlen_k = (
+                    torch.arange(
+                        n_images + 1,
+                        device=device,
+                        dtype=torch.int32,
+                    )
+                    * seqlen_k
+                )
+                max_q = cu_seqlen_q[-1].item()
+                max_k = seqlen_k
+            else:
+                cu_seqlen_q = torch.arange(
+                    seqlen_q + 1, device=device, dtype=torch.int32
+                )
+                seqlen_k = cross_attention_states.shape[1]
+                n_images = cross_attention_states.shape[0]
+                cu_seqlen_k = (
+                    torch.arange(
+                        n_images + 1,
+                        device=device,
+                        dtype=torch.int32,
+                    )
+                    * seqlen_k
+                )
+                max_q = seqlen_q
+                max_k = seqlen_k
+                indices = image_indices[:]
 
-            device = input_ids.device
-            cu_seqlen_q = torch.Tensor([0, seqlen_q]).to(
-                dtype=torch.int32, device=device
-            )
-            cu_seqlen_k = torch.Tensor([0, seqlen_k]).to(
-                dtype=torch.int32, device=device
-            )
-            max_q = seqlen_q
-            max_k = seqlen_k
             cross_attention_states = (
                 cross_attention_states,
                 cu_seqlen_q,
                 cu_seqlen_k,
                 max_q,
                 max_k,
+                indices,
             )
 
         outputs = self.text_model(

server/text_generation_server/models/mllama_causal_lm.py

@@ -1,7 +1,7 @@
 from io import BytesIO
 from PIL import Image
 import torch
-from typing import Iterable
+from typing import Iterable, List
 from text_generation_server.pb.generate_pb2 import Request
 
 from dataclasses import dataclass
@@ -20,10 +20,54 @@ tracer = trace.get_tracer(__name__)
 
 @dataclass
 class MllamaCausalLMBatch(VlmCausalLMBatch):
+    image_indices: List[int] = 42
     aspect_ratio_ids: Optional[torch.Tensor] = None
     aspect_ratio_mask: Optional[torch.Tensor] = None
     cross_attention_states: Optional[torch.Tensor] = None
 
+    @classmethod
+    @tracer.start_as_current_span("concatenate")
+    def concatenate(cls, batches):
+        batch = super().concatenate(batches)
+        batch.pixel_values = None
+        batch.pixel_attention_mask = None
+
+        offset = 0
+        image_indices = []
+        attention_states = []
+        for b in batches:
+            attention_states.append(b.cross_attention_states)
+            image_indices.extend([i + offset for i in b.image_indices])
+            offset += len(b.image_indices)
+        batch.cross_attention_states = torch.cat(attention_states, dim=0)
+        batch.image_indices = image_indices
+        return batch
+
+    @tracer.start_as_current_span("filter")
+    def filter(self, request_ids: List[int]):
+        assert self.image_indices is not None
+        batch = super().filter(request_ids)
+        assert self.image_indices is not None
+        indices = []
+        for i, request_id in enumerate(request_ids):
+            idx = self.requests_idx_mapping[request_id]
+            indices.append(idx)
+
+        offset = 0
+        new_image_indices = []
+        prev_i = None
+        for i in self.image_indices:
+            if i in indices:
+                new_image_indices.append(offset)
+                if i != prev_i:
+                    offset += 1
+                prev_i = i
+
+        batch.image_indices = new_image_indices
+        batch.cross_attention_states = self.cross_attention_states[indices]
+        assert offset <= batch.cross_attention_states.shape[0]
+        return batch
+
     @classmethod
     def batch_tokenized_inputs(
         cls, requests: Iterable[Request], tokenizer, processor, config
@@ -115,5 +159,6 @@ class MllamaCausalLMBatch(VlmCausalLMBatch):
             batch.pixel_values = None
             batch.aspect_ratio_ids = None
             batch.aspect_ratio_mask = None
-            batch.image_indices = None
+            batch.image_indices = []
+        assert batch.image_indices is not None
         return batch

server/text_generation_server/models/vlm_causal_lm.py

@@ -141,7 +141,7 @@ class VlmCausalLMBatch(FlashCausalLMBatch):
     @classmethod
     @tracer.start_as_current_span("concatenate")
     def concatenate(cls, batches):
-        batch = super(VlmCausalLMBatch, cls).concatenate(batches)
+        batch = super().concatenate(batches)
         batch.pixel_values = None
         batch.pixel_attention_mask = None
         batch.image_sizes = None
@@ -402,7 +402,7 @@ class VlmCausalLM(FlashCausalLM):
                     lm_head_indices=lm_head_indices,
                     cross_attention_states=cross_attention_states,
                     adapter_data=adapter_data,
-                    image_indices=batch.image_indices,
+                    image_indices=batch.image_indices[:],
                 )
                 if batch.prefill_cache_indices is not None:
                     batch.prefill_cache_indices = None