v01

.gitattributes

@@ -0,0 +1 @@
+/pytorch_model.bin filter=lfs diff=lfs merge=lfs -text

config.json

@@ -0,0 +1,36 @@
+{
+  "_name_or_path": "/workspace/models/phi-1_5",
+  "activation_function": "gelu_new",
+  "architecture": {
+    "block_cls": "parallel",
+    "mixer": {},
+    "mlp": {
+      "mlp_cls": "mlp"
+    }
+  },
+  "architectures": [
+    "MixFormerSequentialForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_mixformer_sequential.MixFormerSequentialConfig",
+    "AutoModelForCausalLM": "modeling_mixformer_sequential.MixFormerSequentialForCausalLM"
+  },
+  "embd_layer": "default",
+  "embd_pdrop": 0.0,
+  "initializer_range": 0.02,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "mixformer-sequential",
+  "n_embd": 2048,
+  "n_head": 32,
+  "n_inner": null,
+  "n_layer": 24,
+  "n_positions": 2048,
+  "phyagi_version": "0.0.4.dev",
+  "resid_pdrop": 0.0,
+  "rotary_dim": 32,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.33.1",
+  "use_cache": false,
+  "vocab_size": 50304
+}

configuration_mixformer_sequential.py

@@ -0,0 +1,61 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+import math
+from typing import Any, Dict, List, Optional, Union
+
+from transformers import PretrainedConfig
+
+
+class MixFormerSequentialConfig(PretrainedConfig):
+    """MixFormer (sequential for DeepSpeed) configuration."""
+
+    model_type = "mixformer-sequential"
+
+    attribute_map = {
+        "max_position_embeddings": "n_positions",
+        "hidden_size": "n_embd",
+        "num_attention_heads": "n_head",
+        "num_hidden_layers": "n_layer",
+        "input_emb_layer": "embd_layer",  # `input_emb_layer` key is for backward compatibility
+        "blocks": "architecture",  # `blocks` key is for backward compatibility
+    }
+
+    def __init__(
+        self,
+        vocab_size: Optional[int] = 50304,
+        n_positions: Optional[int] = 2048,
+        n_embd: Optional[int] = 1024,
+        n_layer: Optional[int] = 20,
+        n_inner: Optional[int] = None,
+        n_head: Optional[int] = 16,
+        rotary_dim: Optional[int] = 32,
+        activation_function: Optional[str] = "gelu_new",
+        embd_layer: Optional[str] = "default",
+        architecture: Union[Dict[str, Any], List[Dict[str, Any]]] = None,
+        embd_pdrop: Optional[float] = 0.0,
+        resid_pdrop: Optional[float] = 0.0,
+        layer_norm_epsilon: Optional[float] = 1e-5,
+        initializer_range: Optional[float] = 0.02,
+        tie_word_embeddings: Optional[bool] = False,
+        pad_vocab_size_multiple: Optional[int] = 64,
+        **kwargs
+    ) -> None:
+        #self.vocab_size = int(math.ceil(vocab_size / pad_vocab_size_multiple) * pad_vocab_size_multiple)
+        #see https://huggingface.co/teknium/Puffin-Phi/commit/4648d063244250ea9612c241ff996a41b101c9ad
+        self.vocab_size = vocab_size
+        self.n_positions = n_positions
+        self.n_embd = n_embd
+        self.n_layer = n_layer
+        self.n_inner = n_inner
+        self.n_head = n_head
+        self.rotary_dim = min(rotary_dim, n_embd // n_head)
+        self.activation_function = activation_function
+        self.embd_layer = embd_layer
+        self.architecture = architecture
+        self.embd_pdrop = embd_pdrop
+        self.resid_pdrop = resid_pdrop
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+
+        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)

generation_config.json

@@ -0,0 +1,4 @@
+{
+  "_from_model_config": true,
+  "transformers_version": "4.33.1"
+}

latest

@@ -0,0 +1 @@
+global_step2250

modeling_mixformer_sequential.py

@@ -0,0 +1,778 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+# BSD 3-Clause License
+# 
+# Copyright (c) 2022, Tri Dao, trid@cs.stanford.edu.
+# All rights reserved.
+# 
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+# 
+# * Redistributions of source code must retain the above copyright notice, this
+#   list of conditions and the following disclaimer.
+# 
+# * Redistributions in binary form must reproduce the above copyright notice,
+#   this list of conditions and the following disclaimer in the documentation
+#   and/or other materials provided with the distribution.
+# 
+# * Neither the name of the copyright holder nor the names of its
+#   contributors may be used to endorse or promote products derived from
+#   this software without specific prior written permission.
+# 
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+from __future__ import annotations
+
+import math
+import copy
+from typing import Any, Dict, Optional, Tuple
+from dataclasses import dataclass, field
+
+import torch
+import torch.nn as nn
+
+from einops import rearrange
+from transformers.activations import ACT2FN
+from transformers import PretrainedConfig, PreTrainedModel
+from transformers.modeling_outputs import CausalLMOutputWithPast
+
+from .configuration_mixformer_sequential import MixFormerSequentialConfig
+
+@dataclass
+class InferenceParams:
+    """Inference parameters that are passed to the main model in order
+    to efficienly calculate and store the context during inference.
+    Adapted from https://github.com/Dao-AILab/flash-attention."""
+    max_sequence_len: int
+    max_batch_size: int
+    sequence_len_offset: int = 0
+    batch_size_offset: int = 0
+    key_value_memory_dict: dict = field(default_factory=dict)
+    fused_ft_kernel: bool = False
+    lengths_per_sample: Optional[torch.Tensor] = None
+
+
+class Embedding(nn.Module):
+    """Token embedding with dropout."""
+
+    def __init__(self, config: PretrainedConfig) -> None:
+        super().__init__()
+
+        self.wte = nn.Embedding(config.vocab_size, config.n_embd)
+        self.drop = nn.Dropout(config.embd_pdrop)
+
+    def forward(self, input_ids: torch.LongTensor) -> torch.FloatTensor:
+        input_shape = input_ids.size()
+        input_ids = input_ids.view(-1, input_shape[-1])
+
+        hidden_states = self.wte(input_ids)
+        hidden_states = self.drop(hidden_states)
+
+        return hidden_states
+
+class RotaryEmbedding(nn.Module):
+    """PyTorch implementation of `flash-attn` RotaryEmbedding layer.
+    Adapted from https://github.com/Dao-AILab/flash-attention."""
+
+    def __init__(
+        self,
+        dim: int,
+        base: Optional[int] = 10000,
+        scale_base: Optional[float] = None,
+        device: Optional[str] = None,
+        **kwargs,
+    ) -> None:
+        super().__init__()
+
+        if scale_base is not None:
+            raise NotImplementedError
+
+        # Generate and save the inverse frequency buffer (non-trainable)
+        self.dim = dim
+        self.base = base
+        self.scale_base = scale_base
+        self.device = device
+
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, device=device, dtype=torch.float32) / dim))
+        self.register_buffer("inv_freq", inv_freq)
+
+        scale = (
+            (torch.arange(0, dim, 2, device=device, dtype=torch.float32) + 0.4 * dim) / (1.4 * dim)
+            if scale_base is not None
+            else None
+        )
+        self.register_buffer("scale", scale)
+
+        self._seq_len_cached = 0
+        self._cos_cached = None
+        self._sin_cached = None
+        self._cos_k_cached = None
+        self._sin_k_cached = None
+
+    def _update_cos_sin_cache(self, x: torch.FloatTensor, seqlen_offset: Optional[int] = 0) -> None:
+        # Reset the tables if the sequence length has changed,
+        # or if we're on a new device (possibly due to tracing for instance)
+        seqlen = x.shape[1] + seqlen_offset
+
+        # Re-generate the inverse frequency buffer if it's not fp32
+        # (for instance if model.half() was called)
+        if self.inv_freq.dtype != "torch.float32":
+            self.inv_freq = 1.0 / (
+                self.base ** (torch.arange(0, self.dim, 2, device=self.device, dtype=torch.float32) / self.dim)
+            )
+
+        if seqlen > self._seq_len_cached or self._cos_cached.device != x.device or self._cos_cached.dtype != x.dtype:
+            self._seq_len_cached = seqlen
+            t = torch.arange(seqlen, device=x.device, dtype=torch.float32)
+
+            # Don't do einsum, it converts fp32 to fp16
+            # freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+            freqs = torch.outer(t, self.inv_freq.to(device=t.device, dtype=torch.float32))
+            if self.scale is None:
+                self._cos_cached = torch.cos(freqs).to(x.dtype)
+                self._sin_cached = torch.sin(freqs).to(x.dtype)
+            else:
+                power = (
+                    torch.arange(seqlen, dtype=self.scale.dtype, device=self.scale.device) - seqlen // 2
+                ) / self.scale_base
+                scale = self.scale.to(device=power.device) ** rearrange(power, "s -> s 1")
+
+                # We want the multiplication by scale to happen in fp32
+                self._cos_cached = (torch.cos(freqs) * scale).to(x.dtype)
+                self._sin_cached = (torch.sin(freqs) * scale).to(x.dtype)
+                self._cos_k_cached = (torch.cos(freqs) / scale).to(x.dtype)
+                self._sin_k_cached = (torch.sin(freqs) / scale).to(x.dtype)
+
+    def apply_rotary_emb_qkv(
+        self,
+        qkv: torch.FloatTensor,
+        sin: torch.FloatTensor,
+        cos: torch.FloatTensor,
+        sin_k: Optional[torch.FloatTensor] = None,
+        cos_k: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        _, seqlen, three, _, headdim = qkv.shape
+        assert three == 3
+
+        rotary_seqlen, rotary_dim = cos.shape
+        rotary_dim *= 2
+        assert rotary_dim <= headdim
+        assert seqlen <= rotary_seqlen
+
+        cos_k = cos if cos_k is None else cos_k
+        sin_k = sin if sin_k is None else sin_k
+        assert sin.shape == cos_k.shape == sin_k.shape == (rotary_seqlen, rotary_dim // 2)
+
+        q_rot = qkv[:, :, 0, :, :rotary_dim]
+        q_pass = qkv[:, :, 0, :, rotary_dim:]
+
+        k_rot = qkv[:, :, 1, :, :rotary_dim]
+        k_pass = qkv[:, :, 1, :, rotary_dim:]
+
+        # Splits the queries and keys in half
+        q1, q2 = q_rot.chunk(2, dim=-1)
+        k1, k2 = k_rot.chunk(2, dim=-1)
+        c, s = rearrange(cos[:seqlen], "s d -> s 1 d"), rearrange(sin[:seqlen], "s d -> s 1 d")
+
+        # Casts to fp32 are necessary to prevent fp16 overflow issues
+        q1, q2, k1, k2, c, s = [t.to(dtype=torch.float32) for t in [q1, q2, k1, k2, c, s]]
+
+        # Computes the new keys and queries, recasting to original dtype
+        q_rot = torch.cat([q1 * c - q2 * s, q1 * s + q2 * c], axis=-1).to(qkv.dtype)
+
+        k_rot = torch.cat([k1 * c - k2 * s, k1 * s + k2 * c], axis=-1).to(qkv.dtype)
+
+        return torch.cat(
+            [
+                torch.cat([q_rot, q_pass], axis=-1).unsqueeze(2),
+                torch.cat([k_rot, k_pass], axis=-1).unsqueeze(2),
+                qkv[:, :, 2:3, :, :],
+            ],
+            axis=2,
+        )
+
+    def forward(self, qkv: torch.Tensor, seqlen_offset: int = 0) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Perform the forward pass.
+
+        Args:
+            qkv: Query, key and value tensors of shape (batch, seqlen, nheads, headdim) or (batch, seqlen, 3, nheads, headdim).
+            seqlen_offset: Used in generation where the passed `qkv` is only the last token in the batch.
+
+        Returns:
+            New `qkv` and the cached sinusoids.
+
+        """
+
+        self._update_cos_sin_cache(qkv, seqlen_offset)
+
+        return self.apply_rotary_emb_qkv(qkv, self._sin_cached[seqlen_offset:], self._cos_cached[seqlen_offset:])
+
+def _update_kv_cache(kv, inference_params, layer_idx):
+    """kv: (batch_size, seqlen, 2, nheads, head_dim) or (batch_size, 1, 2, nheads, head_dim)
+    Adapted from https://github.com/Dao-AILab/flash-attention."""
+    # Pre-allocate memory for key-values for inference.
+    num_heads, head_dim = kv.shape[-2:]
+    if layer_idx not in inference_params.key_value_memory_dict:
+        kv_cache = torch.empty(
+            inference_params.max_batch_size, inference_params.max_sequence_len, 2,
+            num_heads, head_dim, dtype=kv.dtype, device=kv.device
+        )
+        inference_params.key_value_memory_dict[layer_idx] = kv_cache
+    else:
+        kv_cache = inference_params.key_value_memory_dict[layer_idx]
+
+    # Adjust key and value for inference
+    batch_start = inference_params.batch_size_offset
+    batch_end = batch_start + kv.shape[0]
+    sequence_start = inference_params.sequence_len_offset
+    sequence_end = sequence_start + kv.shape[1]
+    assert batch_end <= (kv_cache.shape[0] if kv_cache is not None else v_cache.shape[0])
+    assert sequence_end <= (kv_cache.shape[1] if kv_cache is not None else v_cache.shape[2])
+
+    assert kv_cache is not None
+    kv_cache[batch_start:batch_end, sequence_start:sequence_end, ...] = kv
+    kv = kv_cache[batch_start:batch_end, :sequence_end, ...]
+    return kv
+
+
+class MLP(nn.Module):
+    """Multi-Layer Perceptron.
+
+    Reference:
+        Attention Is All You Need.
+        https://arxiv.org/pdf/1706.03762.pdf.
+
+    """
+
+    def __init__(self, config: PretrainedConfig, n_inner: Optional[int] = None, act_fn: Optional[str] = None) -> None:
+        super().__init__()
+
+        act_fn = config.activation_function if act_fn is None else act_fn
+        assert act_fn in ACT2FN.keys(), f"`act_fn` must be one of: {ACT2FN.keys()}."
+
+        n_inner = getattr(config, "n_inner", None) if n_inner is None else n_inner
+        n_inner = n_inner if n_inner is not None else 4 * config.n_embd
+
+        self.fc1 = nn.Linear(config.n_embd, n_inner)
+        self.fc2 = nn.Linear(n_inner, config.n_embd)
+        self.act = ACT2FN[act_fn]
+
+    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs):
+        old_keys = [prefix + "fc_in.weight", prefix + "fc_out.weight", prefix + "fc_in.bias", prefix + "fc_out.bias"]
+        new_keys = [prefix + "fc1.weight", prefix + "fc2.weight", prefix + "fc1.bias", prefix + "fc2.bias"]
+        
+        if all(k in state_dict for k in old_keys) and not all(k in state_dict for k in new_keys):
+            # Older version of `MLP` saved with different key names.
+            for old_key, new_key in zip(old_keys, new_keys):
+                state_dict[new_key] = state_dict.pop(old_key)
+
+        return super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs)
+    
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+
+        return hidden_states
+
+
+class FusedMLP(nn.Module):
+    """Fused Multi-Layer Perceptron from `flash-attn`.
+
+    Reference:
+        https://github.com/HazyResearch/flash-attention/blob/main/flash_attn/ops/fused_dense.py.
+
+    """
+    def __init__(self, config: PretrainedConfig, n_inner: Optional[int] = None, act_fn: Optional[str] = None, 
+                 raise_on_missing: bool = False) -> None:
+        super().__init__()
+
+        act_fn = config.activation_function if act_fn is None else act_fn
+        assert act_fn in ACT2FN.keys(), f"`act_fn` must be one of: {ACT2FN.keys()}."
+
+        n_inner = getattr(config, "n_inner", None) if n_inner is None else n_inner
+        n_inner = n_inner if n_inner is not None else 4 * config.n_embd
+
+        gelu_activations = ["gelu_new", "gelu_fast", "gelu_approx"]
+        activation = "gelu_approx" if act_fn in gelu_activations else "relu"
+           
+        self.mlp = MLP(config, n_inner=n_inner, act_fn=act_fn)
+    
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+        return self.mlp(hidden_states)
+
+class SelfAttention(nn.Module):
+    """Implement the scaled dot product attention with softmax.
+    Adapted from https://github.com/Dao-AILab/flash-attention.
+    Arguments
+    ---------
+        softmax_scale: The temperature to use for the softmax attention.
+                      (default: 1/sqrt(d_keys) where d_keys is computed at
+                      runtime)
+        attention_dropout: The dropout rate to apply to the attention
+                           (default: 0.0)
+    """
+    def __init__(self, causal=False, softmax_scale=None, attention_dropout=0.0):
+        super().__init__()
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+        self.drop = nn.Dropout(attention_dropout)
+
+    def forward(self, qkv, causal=None, key_padding_mask=None):
+        """Implements the multihead softmax attention.
+        Arguments
+        ---------
+            qkv: The tensor containing the query, key, and value. (B, S, 3, H, D)
+            causal: if passed, will override self.causal
+            key_padding_mask: boolean mask to apply to the attention weights. True means to keep,
+                False means to mask out. (B, S)
+        """
+        batch_size, seqlen = qkv.shape[0], qkv.shape[1]
+        causal = self.causal if causal is None else causal
+        q, k, v = qkv.unbind(dim=2)
+        softmax_scale = self.softmax_scale or 1.0 / math.sqrt(q.shape[-1])
+        scores = torch.einsum('bthd,bshd->bhts', q, k * softmax_scale)
+        if key_padding_mask is not None:
+            padding_mask = torch.full((batch_size, seqlen), -10000.0, dtype=scores.dtype,
+                                      device=scores.device)
+            padding_mask.masked_fill_(key_padding_mask, 0.0)
+            # TD [2022-09-30]: Adding is faster than masked_fill_ (idk why, just better kernel I guess)
+            scores = scores + rearrange(padding_mask, 'b s -> b 1 1 s')
+        if causal:
+            # "triu_tril_cuda_template" not implemented for 'BFloat16'
+            # So we have to construct the mask in float
+            causal_mask = torch.triu(torch.full((seqlen, seqlen), -10000.0, device=scores.device), 1)
+            # TD [2022-09-30]: Adding is faster than masked_fill_ (idk why, just better kernel I guess)
+            scores = scores + causal_mask.to(dtype=scores.dtype)
+        attention = torch.softmax(scores, dim=-1, dtype=v.dtype)
+        attention_drop = self.drop(attention)
+        output = torch.einsum('bhts,bshd->bthd', attention_drop, v)
+        return output
+
+
+class CrossAttention(nn.Module):
+    """Implement the scaled dot product attention with softmax.
+    Adapted from https://github.com/Dao-AILab/flash-attention.
+    Arguments
+    ---------
+        softmax_scale: The temperature to use for the softmax attention.
+                      (default: 1/sqrt(d_keys) where d_keys is computed at
+                      runtime)
+        attention_dropout: The dropout rate to apply to the attention
+                           (default: 0.0)
+    """
+    def __init__(self, causal=False, softmax_scale=None, attention_dropout=0.0):
+        super().__init__()
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+        self.drop = nn.Dropout(attention_dropout)
+
+    def forward(self, q, kv, causal=None, key_padding_mask=None):
+        """Implements the multihead softmax attention.
+        Arguments
+        ---------
+            q: The tensor containing the query. (B, Sq, H, D)
+            kv: The tensor containing the key and value. (B, Sk, 2, H, D)
+            causal: if passed, will override self.causal
+            key_padding_mask: boolean mask to apply to the attention weights. True means to keep,
+                False means to mask out. (B, Sk)
+        """
+        batch_size, seqlen_q = q.shape[0], q.shape[1]
+        causal = self.causal if causal is None else causal
+        seqlen_k = kv.shape[1]
+        assert kv.shape[0] == batch_size and kv.shape[3] == q.shape[2] and kv.shape[4] == q.shape[3]
+        k, v = kv.unbind(dim=2)
+        softmax_scale = self.softmax_scale or 1.0 / math.sqrt(q.shape[-1])
+        scores = torch.einsum('bthd,bshd->bhts', q, k * softmax_scale)
+        if key_padding_mask is not None:
+            padding_mask = torch.full((batch_size, seqlen_k), -10000.0, dtype=scores.dtype,
+                                      device=scores.device)
+            padding_mask.masked_fill_(key_padding_mask, 0.0)
+            # TD [2022-09-30]: Adding is faster than masked_fill_ (idk why, just better kernel I guess)
+            scores = scores + rearrange(padding_mask, 'b s -> b 1 1 s')
+        if causal:
+            # "triu_tril_cuda_template" not implemented for 'BFloat16'
+            # So we have to construct the mask in float
+            causal_mask = torch.triu(torch.full((seqlen_q, seqlen_k), -10000.0,
+                                                device=scores.device), 1)
+            # TD [2022-09-30]: Adding is faster than masked_fill_ (idk why, just better kernel I guess)
+            scores = scores + causal_mask.to(dtype=scores.dtype)
+        attention = torch.softmax(scores, dim=-1, dtype=v.dtype)
+        attention_drop = self.drop(attention)
+        output = torch.einsum('bhts,bshd->bthd', attention_drop, v)
+        return output
+
+def find_mha_dims(
+    config: PretrainedConfig, n_head: Optional[int] = None, head_dim: Optional[int] = None
+) -> Tuple[int, int]:
+    """Validate and return the number of heads and head dimension for multi-head attention.
+
+    Args:
+        config: Model configuration.
+        n_head: Number of heads.
+        head_dim: Head dimension.
+
+    Returns:
+        Number of heads and head dimension.
+
+    """
+
+    assert all(
+        hasattr(config, attr) for attr in ["n_embd", "n_head"]
+    ), "`config` must have `n_embd` and `n_head` attributes."
+
+    if head_dim is None:
+        assert (
+            config.n_embd % config.n_head == 0
+        ), f"Hidden size ({config.n_embd}) must be divisible by the number of heads ({config.n_head})."
+
+    if n_head is None and head_dim is None:
+        head_dim = config.n_embd // config.n_head
+        n_head = config.n_head
+    elif n_head is None or head_dim is None:
+        raise ValueError("`n_head` and `head_dim` must be both specified or `None`.")
+
+    return n_head, head_dim
+
+
+class MHA(nn.Module):
+    """Multi-head attention layer.
+    Adapted from https://github.com/Dao-AILab/flash-attention."""
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        rotary_dim: Optional[int] = None,
+        n_head: Optional[int] = None,
+        head_dim: Optional[int] = None,
+        bias: Optional[bool] = True,
+        dropout: Optional[float] = 0.0,
+        softmax_scale: Optional[float] = None,
+        causal: Optional[bool] = True,
+        layer_idx: Optional[int] = None,
+        rotary_emb_scale_base: Optional[float] = None,
+        return_residual: Optional[bool] = False,
+        checkpointing: Optional[bool] = False,
+        device: Optional[str] = None,
+        dtype: Optional[torch.dtype] = None,
+        fused_dense: Optional[bool] = True,
+        flash_attn: Optional[bool] = True,
+        cutlass_attn: Optional[bool] = False,
+        flash_rotary: Optional[bool] = True,
+        raise_on_missing: Optional[bool] = False
+    ) -> None:
+        super().__init__()
+
+        factory_kwargs = {"device": device, "dtype": dtype}
+        n_head, head_dim = find_mha_dims(config, n_head, head_dim)
+
+        self.hidden_size = config.n_embd
+        self.n_head = n_head
+        self.head_dim = head_dim
+        self.op_size = n_head * head_dim
+
+        self.causal = causal
+        self.layer_idx = layer_idx
+        self.rotary_emb_dim = rotary_dim if rotary_dim is not None else getattr(config, "rotary_dim", 0)
+        self.fused_dense = fused_dense
+        self.flash_attn = flash_attn
+        self.cutlass_attn = cutlass_attn
+        self.flash_rotary = flash_rotary
+        self.return_residual = return_residual
+        self.checkpointing = checkpointing
+
+        if self.rotary_emb_dim > 0:
+            rotary_kwargs = {"device": device}
+            if rotary_emb_scale_base is not None and rotary_emb_scale_base > 0.0:
+                rotary_kwargs["scale_base"] = rotary_emb_scale_base
+            
+            self.rotary_emb = RotaryEmbedding(self.rotary_emb_dim, **rotary_kwargs)
+        else:
+            pass
+
+        self.Wqkv = nn.Linear(self.hidden_size, 3 * self.op_size, bias=bias, **factory_kwargs)
+        self.out_proj = nn.Linear(self.op_size, self.hidden_size, bias=bias, **factory_kwargs)
+
+        self.inner_attn = SelfAttention(causal=causal, softmax_scale=softmax_scale, attention_dropout=dropout)
+        self.inner_cross_attn = CrossAttention(causal=causal, softmax_scale=softmax_scale, attention_dropout=dropout)
+
+    def _update_kv_cache(self, kv: torch.FloatTensor, inference_params: InferenceParams) -> None:
+        """kv: (batch_size, seqlen, 2, nheads, head_dim) or (batch_size, 1, 2, nheads, head_dim)
+        Adapted from https://github.com/Dao-AILab/flash-attention."""
+
+        assert self.layer_idx is not None, "Generation requires layer_idx in the constructor"
+
+        return _update_kv_cache(kv, inference_params, self.layer_idx)
+
+    def forward(
+        self,
+        x: torch.FloatTensor,
+        x_kv: Optional[torch.FloatTensor] = None,
+        key_padding_mask: Optional[torch.BoolTensor] = None,
+        cu_seqlens: Optional[torch.LongTensor] = None,
+        max_seqlen: Optional[int] = None,
+        mixer_subset: Optional[torch.LongTensor] = None,
+        past_cache: Optional[InferenceParams] = None,
+        **kwargs
+    ) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
+        """Perform the forward pass.
+
+        Args:
+            x: (batch, seqlen, hidden_dim) (where hidden_dim = num heads * head dim) if
+                cu_seqlens is None and max_seqlen is None, else (total, hidden_dim) where total
+                is the is the sum of the sequence lengths in the batch.
+            x_kv: (batch, seqlen, hidden_dim), only applicable for cross-attention. If None, use x.
+            key_padding_mask: boolean mask, True means to keep, False means to mask out.
+                (batch, seqlen). Only applicable when not using FlashAttention.
+            cu_seqlens: (batch_size + 1,), dtype torch.int32. The cumulative sequence lengths
+                of the sequences in the batch, used to index into x. Only applicable when using
+                FlashAttention.
+            max_seqlen: int. Maximum sequence length in the batch.
+            mixer_subset: for cross-attention only. If not None, will take a subset of x
+                before applying the query projection. Useful for e.g., ViT where we only care
+                about the CLS token in the last layer.
+            past_cache: For generation only.
+
+        Returns:
+            (batch, seqlen, hidden_dim) if cu_seqlens is None and max_seqlen is None,
+                else (total, hidden_dim) where total is the is the sum of the sequence lengths
+                in the batch.
+
+        """
+
+        if cu_seqlens is not None:
+            assert max_seqlen is not None
+            assert key_padding_mask is None
+            assert self.flash_attn
+            assert self.rotary_emb_dim == 0
+
+        if key_padding_mask is not None:
+            assert cu_seqlens is None
+            assert max_seqlen is None
+            assert not self.flash_attn
+
+        if past_cache is not None:
+            assert key_padding_mask is None
+            assert cu_seqlens is None and max_seqlen is None
+
+        attn_kwargs = {"key_padding_mask": key_padding_mask}
+
+        assert x_kv is None and mixer_subset is None
+
+        qkv = self.Wqkv(x)
+        qkv = rearrange(qkv, "... (three h d) -> ... three h d", three=3, d=self.head_dim)
+
+        if past_cache is None:
+            if self.rotary_emb_dim > 0:
+                qkv = self.rotary_emb(qkv)
+                context = self.inner_attn(qkv, **attn_kwargs)
+
+        else:
+            if self.rotary_emb_dim > 0:
+                qkv = self.rotary_emb(qkv, seqlen_offset=past_cache.sequence_len_offset)
+            q = qkv[:, :, 0]
+            kv = self._update_kv_cache(qkv[:, :, 1:], past_cache)
+            # If we're processing the prompt, causal=None (use self.causal).
+            # If we're decoding, then causal=False.
+            causal = None if past_cache.sequence_len_offset == 0 else False
+            context = self.inner_cross_attn(q, kv, causal=causal)
+
+        out = rearrange(context, "... h d -> ... (h d)")
+        out = self.out_proj(out)
+
+        return out if not self.return_residual else (out, x)
+
+class ParallelBlock(nn.Module):
+    """Parallel block.
+
+    This block applies parallel mixer and MLP layers to the input (used in GPT-J and CodeGen).
+
+    """
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        mixer: Optional[Dict[str, Any]] = None,
+        mlp: Optional[Dict[str, Any]] = None,
+        block_idx: Optional[int] = None,
+    ) -> None:
+        super().__init__()
+
+        self.ln = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        self.resid_dropout = nn.Dropout(config.resid_pdrop)
+        self.block_idx = block_idx
+
+        self.mixer = MHA(config=config, **mixer, layer_idx=block_idx)
+        mlp_cls = mlp.pop('mlp_cls')
+        if mlp_cls == 'fused_mlp':
+            self.mlp = FusedMLP(config=config, **mlp)
+        else:
+            self.mlp = MLP(config=config, **mlp)
+
+    def forward(self, hidden_states: torch.FloatTensor, 
+                past_cache: Optional[torch.FloatTensor] = None) -> torch.FloatTensor:
+        residual = hidden_states
+        hidden_states = self.ln(hidden_states)
+
+        attn_outputs = self.mixer(hidden_states, past_cache=past_cache)
+        if isinstance(attn_outputs, tuple):
+            attn_outputs = attn_outputs[0]
+
+        attn_outputs = self.resid_dropout(attn_outputs)
+        feed_forward_hidden_states = self.resid_dropout(self.mlp(hidden_states))
+
+        hidden_states = attn_outputs + feed_forward_hidden_states + residual
+
+        return hidden_states
+
+class CausalLMHead(nn.Module):
+    """Causal Language Modeling head.
+
+    Reference:
+        Improving Language Understanding by Generative Pre-Training.
+        https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf.
+
+    """
+
+    def __init__(self, config: PretrainedConfig) -> None:
+        super().__init__()
+
+        self.ln = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        self.linear = nn.Linear(config.n_embd, config.vocab_size)
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+        hidden_states = self.ln(hidden_states)
+        logits = self.linear(hidden_states).to(torch.float32)
+
+        return logits
+
+
+class CausalLMLoss(nn.Module):
+    """Causal Language Modeling loss.
+
+    Reference:
+        Improving Language Understanding by Generative Pre-Training.
+        https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf.
+
+    """
+
+    def __init__(self, shift_labels: Optional[bool] = True) -> None:
+        super().__init__()
+
+        self.shift_labels = shift_labels
+        self.loss_fct = nn.CrossEntropyLoss()
+
+    def forward(self, logits: torch.FloatTensor, labels: torch.LongTensor) -> torch.FloatTensor:
+        if self.shift_labels:
+            logits = logits[..., :-1, :].contiguous()
+            labels = labels[..., 1:].contiguous()
+
+        loss = self.loss_fct(logits.view(-1, logits.size(-1)), labels.view(-1))
+
+        return loss
+
+class MixFormerSequentialPreTrainedModel(PreTrainedModel):
+    """MixFormer (sequential for DeepSpeed) pre-trained model."""
+
+    config_class = MixFormerSequentialConfig
+    base_model_prefix = "transformer"
+    supports_gradient_checkpointing = True
+
+    def __init__(self, *inputs, **kwargs) -> None:
+        super().__init__(*inputs, **kwargs)
+
+    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, **kwargs) -> Dict[str, Any]:
+        if "use_cache" in kwargs and not kwargs["use_cache"]:
+            return {"input_ids": input_ids}
+
+        if past_key_values is None or not (isinstance(past_key_values, InferenceParams)):
+            past_key_values = InferenceParams(
+                max_batch_size=input_ids.shape[0],
+                max_sequence_len=self.config.n_positions,
+                sequence_len_offset=0,
+                batch_size_offset=0,
+                fused_ft_kernel=False,
+                key_value_memory_dict={},
+            )
+        else:
+            # assume past_key_values has cached all but last token in input_ids
+            past_key_values.sequence_len_offset = len(input_ids[0]) - 1
+            input_ids = input_ids[:, -1].unsqueeze(-1)
+
+        return {"input_ids": input_ids, "past_key_values": past_key_values, **kwargs}
+
+
+class MixFormerSequentialForCausalLM(MixFormerSequentialPreTrainedModel):
+    """MixFormer (sequential for DeepSpeed) for Causal Language Modeling."""
+
+    _keys_to_ignore_on_load_missing = [""]
+    _keys_to_ignore_on_load_unexpected = [r"layers\.\d+\.mlp.(fc_in|fc_out)\.(weight|bias)"]
+
+    def __init__(self, config: MixFormerSequentialConfig) -> None:
+        super().__init__(config)
+
+        modules = [Embedding(config)]
+        block_config = config.architecture
+
+        if not isinstance(block_config, list):
+            block_config = [block_config for _ in range(config.n_layer)]
+
+        if config.n_layer != len(block_config):
+            config.n_layer = len(block_config)
+
+        for block_idx, block in enumerate(block_config):
+            # `block_cls` with `legacy` value is for backward compatibility
+            # `path` key is for backward compatibility
+            block = copy.deepcopy(block) or {"block_cls": "parallel"}
+            block_cls = block.pop("path", None) or block.pop("block_cls", None)
+
+            block["block_idx"] = block_idx
+            modules.append(ParallelBlock(config, **block))
+
+        modules.append(CausalLMHead(config))
+
+        self.layers = nn.Sequential(*modules)
+        self.loss = CausalLMLoss()
+
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.layers[0].wte
+
+    def set_input_embeddings(self, new_embeddings: nn.Embedding) -> None:
+        self.layers[0].wte = new_embeddings
+
+    def get_output_embeddings(self) -> nn.Linear:
+        return self.layers[-1].linear
+
+    def set_output_embeddings(self, new_embeddings: nn.Linear) -> None:
+        self.layers[-1].linear = new_embeddings
+
+    def forward(
+        self, input_ids: torch.LongTensor, labels: Optional[torch.LongTensor] = None, 
+        past_key_values: Optional[torch.FloatTensor] = None, **kwargs
+    ) -> CausalLMOutputWithPast:
+
+        if not past_key_values:
+            lm_logits = self.layers(input_ids)
+        else:
+            hidden_layer = self.layers[0](input_ids)
+            for module in self.layers[1:-1]:
+                hidden_layer = module(hidden_layer, past_cache=past_key_values)
+            lm_logits = self.layers[-1](hidden_layer)
+
+        loss = None
+        if labels is not None:
+            loss = self.loss(lm_logits, labels)
+        
+        return CausalLMOutputWithPast(loss=loss, logits=lm_logits, past_key_values=past_key_values)

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:853d337655539907af7065c571a1c84bc70cdc932182625223df61ca2c13804b
+size 2829283838

zero_to_fp32.py

@@ -0,0 +1,587 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dict = torch.load(f, map_location=device)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage <= 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel()
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)