Revamp medusa implementation so that every model can benefit. (#1588)

# What does this PR do?

<!--
Congratulations! You've made it this far! You're not quite done yet
though.

Once merged, your PR is going to appear in the release notes with the
title you set, so make sure it's a great title that fully reflects the
extent of your awesome contribution.

Then, please replace this with a description of the change and which
issue is fixed (if applicable). Please also include relevant motivation
and context. List any dependencies (if any) that are required for this
change.

Once you're done, someone will review your PR shortly (see the section
"Who can review?" below to tag some potential reviewers). They may
suggest changes to make the code even better. If no one reviewed your PR
after a week has passed, don't hesitate to post a new comment
@-mentioning the same persons---sometimes notifications get lost.
-->

<!-- Remove if not applicable -->

Fixes # (issue)


## Before submitting
- [ ] This PR fixes a typo or improves the docs (you can dismiss the
other checks if that's the case).
- [ ] Did you read the [contributor
guideline](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#start-contributing-pull-requests),
      Pull Request section?
- [ ] Was this discussed/approved via a Github issue or the
[forum](https://discuss.huggingface.co/)? Please add a link
      to it if that's the case.
- [ ] Did you make sure to update the documentation with your changes?
Here are the
[documentation
guidelines](https://github.com/huggingface/transformers/tree/main/docs),
and
[here are tips on formatting
docstrings](https://github.com/huggingface/transformers/tree/main/docs#writing-source-documentation).
- [ ] Did you write any new necessary tests?


## Who can review?

Anyone in the community is free to review the PR once the tests have
passed. Feel free to tag
members/contributors who may be interested in your PR.

<!-- Your PR will be replied to more quickly if you can figure out the
right person to tag with @


@OlivierDehaene OR @Narsil

 -->

server/text_generation_server/utils/hub.py

@@ -40,6 +40,7 @@ def _weight_hub_files_from_model_info(
         and "arguments" not in s.rfilename
         and "args" not in s.rfilename
         and "training" not in s.rfilename
+        and "medusa_lm_head" not in s.rfilename
     ]
 
 
@@ -56,6 +57,7 @@ def _weight_files_from_dir(d: Path, extension: str) -> List[str]:
         and "args" not in f
         and "adapter" not in f
         and "training" not in f
+        and "medusa_lm_head" not in f
     ]
     return filenames
 

server/text_generation_server/utils/layers.py

@@ -4,7 +4,7 @@ import torch.distributed
 
 from torch import nn
 from torch.nn import functional as F
-from typing import List
+from typing import List, Tuple, Optional
 from loguru import logger
 from functools import lru_cache
 
@@ -380,6 +380,96 @@ class SuperLayer(nn.Module):
         return self.linear.forward(x)
 
 
+class ResBlock(torch.nn.Module):
+    def __init__(self, config, prefix, weights):
+        super().__init__()
+        self.linear = FastLinear.load(
+            config, prefix=f"{prefix}.linear", weights=weights, bias=True
+        )
+        self.act = torch.nn.SiLU()
+
+    def forward(self, x):
+        return x + self.act(self.linear(x))
+
+
+class MedusaModel(torch.nn.Module):
+    def __init__(self, config, weights):
+        super().__init__()
+        self.heads = torch.nn.ModuleList(
+            [
+                MedusaHead(config, prefix=f"{i}", weights=weights)
+                for i in range(config["medusa_num_heads"])
+            ]
+        )
+
+    def forward(self, x):
+        speculative_logits = torch.stack([head(x) for head in self.heads], dim=1)
+        return speculative_logits
+
+
+class MedusaHead(torch.nn.Module):
+    def __init__(self, config, prefix, weights):
+        super().__init__()
+        self.blocks = torch.nn.ModuleList(
+            [
+                ResBlock(config, prefix=f"{prefix}.{i}", weights=weights)
+                for i in range(config["medusa_num_layers"])
+            ]
+        )
+        n = len(self.blocks)
+        self.out = FastLinear.load(
+            config, prefix=f"{prefix}.{n}", weights=weights, bias=False
+        )
+
+    def forward(self, x):
+        for block in self.blocks:
+            x = block(x)
+        x = self.out(x)
+        return x
+
+
+class SpeculativeHead(nn.Module):
+    def __init__(self, lm_head, medusa):
+        super().__init__()
+        self.lm_head = lm_head
+        self.medusa = medusa
+
+    @staticmethod
+    def load(config, prefix: str, weights):
+        lm_head = TensorParallelHead.load(config, prefix, weights)
+        use_medusa = config.use_medusa
+        if use_medusa:
+            from pathlib import Path
+            from safetensors import safe_open
+            import json
+
+            medusa_config = str(Path(use_medusa) / "config.json")
+            filename = str(Path(use_medusa) / "medusa_lm_head.safetensors")
+
+            with open(medusa_config, "r") as f:
+                config = json.load(f)
+            routing = weights.routing
+            with safe_open(filename, framework="pytorch") as f:
+                for k in f.keys():
+                    if k in routing:
+                        raise RuntimeError(
+                            f"Key {k} was found in multiple files: {filename} and {routing[k]}"
+                        )
+                    weights.routing[k] = filename
+
+            medusa = MedusaModel(config, weights)
+        else:
+            medusa = None
+        return SpeculativeHead(lm_head, medusa)
+
+    def forward(
+        self, input: torch.Tensor
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        logits = self.lm_head(input)
+        speculative_logits = self.medusa(input) if self.medusa is not None else None
+        return logits, speculative_logits
+
+
 class TensorParallelHead(SuperLayer):
     def __init__(self, linear, process_group, should_gather: bool):
         super().__init__(linear)

server/text_generation_server/utils/medusa.py

-import torch
-from dataclasses import dataclass
-from text_generation_server.utils.layers import TensorParallelHead, FastLinear
-
-
-@dataclass
-class Output:
-    logits: torch.FloatTensor = None
-    speculative_logits: torch.FloatTensor = None
-
-
-class ResBlock(torch.nn.Module):
-    def __init__(self, config, prefix, weights):
-        super().__init__()
-        self.linear = FastLinear.load(
-            config, prefix=f"{prefix}.linear", weights=weights, bias=True
-        )
-        self.act = torch.nn.SiLU()
-
-    def forward(self, x):
-        return x + self.act(self.linear(x))
-
-
-class MedusaModel(torch.nn.Module):
-    def __init__(self, config, weights, lm_head):
-        super().__init__()
-        self.heads = torch.nn.ModuleList(
-            [
-                MedusaHead(config, prefix=f"{i}", weights=weights)
-                for i in range(config["medusa_num_heads"])
-            ]
-        )
-        self.lm_head = lm_head
-
-    def forward(self, x):
-        logits = self.lm_head(x)
-        speculative_logits = torch.stack([head(x) for head in self.heads], dim=1)
-        return logits, speculative_logits
-
-
-class MedusaHead(torch.nn.Module):
-    def __init__(self, config, prefix, weights):
-        super().__init__()
-        self.blocks = torch.nn.ModuleList(
-            [
-                ResBlock(config, prefix=f"{prefix}.{i}", weights=weights)
-                for i in range(config["medusa_num_layers"])
-            ]
-        )
-        n = len(self.blocks)
-        self.out = FastLinear.load(
-            config, prefix=f"{prefix}.{n}", weights=weights, bias=False
-        )
-
-    def forward(self, x):
-        for block in self.blocks:
-            x = block(x)
-        x = self.out(x)
-        return x