feat: load phi weights and produce nonsense tokens

server/text_generation_server/models/__init__.py

@@ -58,6 +58,7 @@ try:
     from text_generation_server.models.idefics import IDEFICSSharded
     from text_generation_server.models.flash_mistral import FlashMistral
     from text_generation_server.models.flash_mixtral import FlashMixtral
+    from text_generation_server.models.flash_phi import FlashPhi
     from text_generation_server.utils.flash_attn import HAS_FLASH_ATTN_V2_CUDA
 
 except ImportError as e:
@@ -73,6 +74,7 @@ if FLASH_ATTENTION:
     __all__.append(IDEFICSSharded)
     __all__.append(FlashMistral)
     __all__.append(FlashMixtral)
+    __all__.append(FlashPhi)
 
 
 def get_model(
@@ -229,11 +231,22 @@ def get_model(
                 trust_remote_code=trust_remote_code,
             )
         
+    elif model_type == "phi":
+        if FLASH_ATTENTION:
+            return FlashPhi(
+                model_id,
+                revision,
+                quantize=quantize,
+                dtype=dtype,
+                trust_remote_code=trust_remote_code,
+                use_medusa=use_medusa,
+            )
+        else:
+            raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Phi"))
+
     elif model_type == "phi-msft":
         if FLASH_ATTENTION:
-            raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Flash Phi"))
+            raise NotImplementedError("Legacy phi-msft is not supported with Flash Attention")
-        elif sharded:
-            raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Sharded Phi"))
         else:
             return Phi(
                 model_id,

server/text_generation_server/models/custom_modeling/flash_phi_modeling.py

@@ -0,0 +1,435 @@
+import torch
+import torch.distributed
+
+from torch import nn
+from transformers.activations import ACT2FN
+from transformers.configuration_utils import PretrainedConfig
+from typing import Optional, List, Tuple
+
+from text_generation_server.utils import paged_attention, flash_attn
+from text_generation_server.utils.layers import (
+    TensorParallelRowLinear,
+    TensorParallelColumnLinear,
+    TensorParallelEmbedding,
+    PositionRotaryEmbedding,
+    TensorParallelHead,
+    get_linear,
+    FastRMSNorm,
+    FastLayerNorm,
+    FastLinear,
+)
+
+class PhiConfig(PretrainedConfig):
+    def __init__(
+        self,
+        vocab_size=51200,
+        hidden_size=2560,
+        intermediate_size=11008,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=None,
+        hidden_act="gelu_fast",
+        max_position_embeddings=2048,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        pretraining_tp=1,
+        tie_word_embeddings=False,
+        rope_scaling=None,
+        rope_theta=10000.0,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.pretraining_tp = pretraining_tp
+        self.use_cache = use_cache
+        self.rope_scaling = rope_scaling
+        self.rope_theta = rope_theta
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+
+def load_attention(config, prefix, weights):
+    if config.num_attention_heads != config.num_key_value_heads:
+        # should never get here
+        return _load_gqa(config, prefix, weights)
+    else:
+        if config.model_type == "baichuan":
+            return TensorParallelColumnLinear.load_qkv(
+                config,
+                prefix=f"{prefix}.W_pack",
+                weights=weights,
+                bias=False,
+            )
+        else:
+            # should be here
+            return TensorParallelColumnLinear.load_multi(
+                config,
+                prefixes=[f"{prefix}.q_proj", f"{prefix}.k_proj", f"{prefix}.v_proj"],
+                dim=0,
+                weights=weights,
+                bias=False,
+            )
+
+
+def _load_gqa(config, prefix: str, weights):
+    assert config.hidden_size % config.num_attention_heads == 0
+    assert config.num_attention_heads % weights.process_group.size() == 0
+
+    weight = weights.get_multi_weights_col(
+        prefixes=[f"{prefix}.q_proj", f"{prefix}.k_proj", f"{prefix}.v_proj"],
+        quantize=config.quantize,
+        dim=0,
+    )
+
+    if config.quantize not in ["gptq", "awq"]:
+        weight = weight.to(dtype=weights.dtype).to(device=weights.device)
+
+        head_size = config.hidden_size // config.num_attention_heads
+        num_heads = config.num_attention_heads // weights.process_group.size()
+        num_key_value_heads = config.num_key_value_heads // weights.process_group.size()
+        assert list(weight.shape) == [
+            (num_heads + 2 * num_key_value_heads) * head_size,
+            config.hidden_size,
+        ], f"{list(weight.shape)} != {[(num_heads + 2 * config.num_key_value_heads) * head_size, config.hidden_size]}"
+
+    return TensorParallelColumnLinear(
+        get_linear(weight, bias=None, quantize=config.quantize)
+    )
+
+
+class FlashPhiAttention(torch.nn.Module):
+    def __init__(
+        self,
+        prefix: str,
+        config,
+        weights,
+    ):
+        super().__init__()
+        self.num_heads = config.num_attention_heads
+        self.hidden_size = config.hidden_size
+        self.head_size = self.hidden_size // self.num_heads
+
+        # MAYBE (if not static)
+        self.rotary_emb = PositionRotaryEmbedding.static(
+            config=config,
+            dim=self.head_size,
+            base=config.rope_theta,
+            device=weights.device,
+        )
+
+        self.softmax_scale = self.head_size**-0.5
+
+        if self.num_heads % weights.process_group.size() != 0:
+            raise ValueError(
+                f"`num_heads` must be divisible by `num_shards` (got `num_heads`: {self.num_heads} "
+                f"and `num_shards`: {weights.process_group.size()}"
+            )
+
+        # should be correct
+        self.num_heads = self.num_heads // weights.process_group.size()
+        self.num_key_value_heads = (
+            config.num_key_value_heads // weights.process_group.size()
+        )
+
+        self.query_key_value = load_attention(config, prefix, weights)
+
+        self.dense = TensorParallelRowLinear.load(
+            config,
+            prefix=f"{prefix}.dense",
+            weights=weights,
+            bias=False,
+        )
+        self.num_groups = self.num_heads // self.num_key_value_heads
+        self.kv_head_mapping = torch.arange(
+            0, self.num_key_value_heads, dtype=torch.int32, device=weights.device
+        ).repeat_interleave(self.num_groups)
+
+    def forward(
+        self,
+        hidden_states,
+        cos,
+        sin,
+        cu_seqlen_prefill,
+        kv_cache,
+        block_tables,
+        slots,
+        input_lengths,
+        max_s,
+    ):
+        qkv = self.query_key_value(hidden_states)
+        query, kv = qkv.split(
+            [
+                self.head_size * self.num_heads,
+                2 * self.head_size * self.num_key_value_heads,
+            ],
+            dim=1,
+        )
+        query = query.view(-1, self.num_heads, self.head_size)
+        kv = kv.view(-1, 2, self.num_key_value_heads, self.head_size)
+
+        self.rotary_emb(query, torch.select(kv, dim=1, index=0), cos, sin)
+
+        paged_attention.reshape_and_cache(
+            kv[:, 0], kv[:, 1], kv_cache[0], kv_cache[1], slots
+        )
+
+        # output tensor
+        attn_output = torch.empty_like(query)
+
+        # Prefill
+        if cu_seqlen_prefill is not None:
+            # flash attention
+            flash_attn.attention(
+                query,
+                torch.select(kv, dim=1, index=0),
+                torch.select(kv, dim=1, index=1),
+                attn_output,
+                cu_seqlen_prefill,
+                max_s,
+                self.softmax_scale,
+            )
+        # Decode
+        else:
+            paged_attention.attention(
+                attn_output,
+                query,
+                kv_cache[0],
+                kv_cache[1],
+                self.kv_head_mapping,
+                self.softmax_scale,
+                block_tables,
+                input_lengths,
+                max_s,
+            )
+
+        return self.dense(attn_output.view(-1, self.num_heads * self.head_size))
+
+
+class PhiMLP(nn.Module):
+    def __init__(self, prefix, config, weights):
+        super().__init__()
+        act = config.hidden_act
+        self.act = (
+            ACT2FN[act]
+            if "gelu" not in act
+            else lambda x: torch.nn.functional.gelu(
+                x,
+                approximate="tanh"
+                if act in ["gelu_fast", "gelu_pytorch_tanh"]
+                else "none",
+            )
+        )
+
+        self.gate_up_proj = TensorParallelRowLinear.load(
+            config,
+            prefix=f"{prefix}.fc1",
+            weights=weights,
+            bias=False,
+        )
+        self.down_proj = TensorParallelRowLinear.load(
+            config,
+            prefix=f"{prefix}.fc2",
+            weights=weights,
+            bias=False,
+        )
+
+    def forward(self, hidden_states):
+        gate_up_states = self.gate_up_proj(hidden_states)
+        post_act = self.act(gate_up_states)
+        return self.down_proj(post_act)
+
+
+class FlashPhiLayer(nn.Module):
+    def __init__(self, layer_id, config, weights):
+        super().__init__()
+        prefix = f"model.layers.{layer_id}"
+        self.self_attn = FlashPhiAttention(
+            prefix=f"{prefix}.self_attn", config=config, weights=weights
+        )
+        self.mlp = PhiMLP(prefix=f"{prefix}.mlp", config=config, weights=weights)
+
+        self.input_layernorm = FastRMSNorm.load(
+            prefix=f"{prefix}.input_layernorm", weights=weights, eps=config.rms_norm_eps
+        )
+
+    def forward(
+        self,
+        hidden_states,
+        residual,
+        cos,
+        sin,
+        cu_seqlen_prefill,
+        kv_cache,
+        block_tables,
+        slots,
+        input_lengths,
+        max_s,
+    ):
+        normed_hidden_states, res = self.input_layernorm(hidden_states, residual)
+
+        # Self Attention
+        attn_output = self.self_attn(
+            normed_hidden_states,
+            cos,
+            sin,
+            cu_seqlen_prefill,
+            kv_cache,
+            block_tables,
+            slots,
+            input_lengths,
+            max_s,
+        )
+
+
+        mlp_output = self.mlp(normed_hidden_states)
+
+        result =  attn_output + mlp_output + res
+
+        return result, res
+
+
+class FlashPhiModel(torch.nn.Module):
+    def __init__(self, config, weights):
+        super().__init__()
+
+        process_group = weights.process_group
+        self.tp_rank = process_group.rank()
+        self.tp_world_size = process_group.size()
+        self.embed_tokens = TensorParallelEmbedding(
+            prefix="model.embed_tokens", weights=weights
+        )
+        self.layers = nn.ModuleList(
+            [
+                FlashPhiLayer(
+                    layer_id,
+                    config,
+                    weights,
+                )
+                for layer_id in range(config.num_hidden_layers)
+            ]
+        )
+        self.gradient_checkpointing = False
+
+        self.head_size = self.layers[0].self_attn.head_size
+        self.num_heads = self.layers[0].self_attn.num_heads
+        self.num_key_value_heads = self.layers[0].self_attn.num_key_value_heads
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        position_ids: torch.Tensor,
+        cu_seqlen_prefill: Optional[torch.Tensor],
+        kv_cache: List[Tuple[torch.Tensor, torch.Tensor]],
+        block_tables: torch.Tensor,
+        slots: torch.Tensor,
+        input_lengths: torch.Tensor,
+        max_s: int,
+    ) -> torch.Tensor:
+        hidden_states = self.embed_tokens(input_ids)
+
+        # Get rotary cos and sin for this forward
+        # Avoid to index in each layer
+        cos, sin = self.layers[0].self_attn.rotary_emb.get_cos_sin(
+            position_ids, max_s, hidden_states.dtype
+        )
+
+        residual = None
+        for i, layer in enumerate(self.layers):
+            hidden_states, residual = layer(
+                hidden_states,
+                residual,
+                cos,
+                sin,
+                cu_seqlen_prefill,
+                kv_cache[i],
+                block_tables,
+                slots,
+                input_lengths,
+                max_s,
+            )
+
+        return hidden_states
+
+
+class FlashPhiForCausalLM(torch.nn.Module):
+    def __init__(self, config, weights):
+        super().__init__()
+
+        self.model = FlashPhiModel(config, weights)
+        # self.lm_head = TensorParallelHead.load(
+        #     config,
+        #     prefix="lm_head",
+        #     weights=weights,
+        # )
+
+        # TODO: prefer parallel head
+        self.linear = FastLinear.load(
+            config,
+            prefix="lm_head",
+            weights=weights,
+            bias=False,
+        )
+
+        # TODO: use in correct place
+        self.ln = FastLayerNorm.load(
+            prefix="model.final_layernorm",
+            weights=weights,
+            eps=config.rms_norm_eps,
+        )
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        position_ids: torch.Tensor,
+        # 1000K and 10K 
+        cu_seqlen_prefill: Optional[torch.Tensor], # indexes for the items in the batch
+        kv_cache: List[Tuple[torch.Tensor, torch.Tensor]],
+        # paged attention related 
+        block_tables: torch.Tensor, # <- indexes into blocks 
+        slots: torch.Tensor, # <- indexes into mem
+        input_lengths: torch.Tensor, 
+        # both attentions
+        max_s: int, # <- max sequence length (make kernals chose swap)
+        # small opt (only care about final)
+        lm_head_indices: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        hidden_states = self.model(
+            input_ids,
+            position_ids,
+            cu_seqlen_prefill,
+            kv_cache,
+            block_tables,
+            slots,
+            input_lengths,
+            max_s,
+        )
+        if lm_head_indices is not None:
+            hidden_states = hidden_states[lm_head_indices]
+
+        normed_hidden_states, res = self.ln(hidden_states, None)
+        logits = self.linear(normed_hidden_states)
+
+        return logits

server/text_generation_server/models/flash_phi.py

@@ -0,0 +1,102 @@
+import torch
+import torch.distributed
+
+from opentelemetry import trace
+from transformers import AutoConfig, AutoTokenizer
+from typing import Optional
+
+from text_generation_server.models import FlashCausalLM
+from text_generation_server.models.custom_modeling.flash_phi_modeling import (
+    FlashPhiForCausalLM,
+    PhiConfig,
+)
+from text_generation_server.utils import (
+    initialize_torch_distributed,
+    weight_files,
+    Weights,
+)
+
+tracer = trace.get_tracer(__name__)
+
+
+class FlashPhi(FlashCausalLM):
+    def __init__(
+        self,
+        model_id: str,
+        revision: Optional[str] = None,
+        quantize: Optional[str] = None,
+        dtype: Optional[torch.dtype] = None,
+        trust_remote_code: bool = False,
+        use_medusa: Optional[str] = None,
+    ):
+        self.process_group, rank, world_size = initialize_torch_distributed()
+        if torch.cuda.is_available():
+            device = torch.device(f"cuda:{rank}")
+            dtype = torch.float16 if dtype is None else dtype
+        else:
+            raise NotImplementedError("FlashPhi is only available on GPU")
+
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_id,
+            revision=revision,
+            padding_side="left",
+            truncation_side="left",
+            trust_remote_code=trust_remote_code,
+        )
+
+        config = PhiConfig.from_pretrained(
+            model_id, revision=revision, trust_remote_code=trust_remote_code
+        )
+        config.quantize = quantize
+
+        torch.distributed.barrier(group=self.process_group)
+
+        filenames = weight_files(model_id, revision=revision, extension=".safetensors")
+        weights = Weights(filenames, device, dtype, process_group=self.process_group)
+        if config.quantize in ["gptq", "awq"]:
+            weights._set_gptq_params(model_id, revision)
+
+        model = FlashPhiForCausalLM(config, weights)
+        if use_medusa:
+            from text_generation_server.utils.medusa import MedusaModel
+            from huggingface_hub import hf_hub_download
+            import json
+            import os
+            from pathlib import Path
+            
+            is_local_model = (Path(use_medusa).exists() and Path(use_medusa).is_dir()) or os.getenv(
+                "WEIGHTS_CACHE_OVERRIDE", None
+            ) is not None
+            
+            if not is_local_model:
+                medusa_config = hf_hub_download(
+                    use_medusa, revision=revision, filename="config.json"
+                )
+                medusa_head = hf_hub_download(
+                    use_medusa, revision=revision, filename="medusa_lm_head.pt"
+                )
+            else:
+                medusa_config = str(Path(use_medusa) / "config.json")
+                medusa_head = str(Path(use_medusa) / "medusa_lm_head.pt")
+                
+            with open(medusa_config, "r") as f:
+                config = json.load(f)
+            medusa_sf = medusa_head[: -len(".pt")] + ".safetensors"
+            weights = Weights(
+                [medusa_sf], device, dtype, process_group=self.process_group
+            )
+            lm_head = model.lm_head
+            model.lm_head = MedusaModel(config, weights, lm_head)
+
+        torch.distributed.barrier(group=self.process_group)
+        super(FlashPhi, self).__init__(
+            model=model,
+            tokenizer=tokenizer,
+            num_layers=len(model.model.layers),
+            num_kv_heads=model.model.num_key_value_heads,
+            head_size=model.model.head_size,
+            dtype=dtype,
+            device=device,
+            rank=rank,
+            world_size=world_size,
+        )