feat(server): Using `quantize_config.json` instead of GPTQ_BITS env variables. (#671)

- Current PR is not great because we're side stepping the
  `Weights.__init__` but Weights shouldn't requires anything related
  to the config or the model_id as it aims to be a simple Wrapper
  over multi file loading.
- Ideal solution would be to use something like Rust enum
  ```
  enum Quantize{
    Bitandbytes(Bitsandbytes),
    GPTQ(bits: usize, groupsize: usize)
  ```
  And passing that around during load. Unfortunately we don't
  have access to this, so for now, side-stepping seems easier.

- Re-enabling groupsize<0 with exllama (confirmed it works.)

Helps #601 

In next steps we should make sure our quantization script uses that
format and make it standard.


# 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/models/bloom.py

@@ -76,6 +76,8 @@ class BLOOMSharded(CausalLM):
         weights = Weights(
             filenames, device=device, dtype=dtype, process_group=self.process_group
         )
+        if config.quantize == "gptq":
+            weights._set_gptq_params(model_id)
 
         model = BloomForCausalLM(config, weights)
 

server/text_generation_server/models/custom_modeling/flash_santacoder_modeling.py

@@ -76,7 +76,7 @@ def _load_multi_mqa_gptq(
 
         g_idx = weights.get_tensor(f"{prefix}.c_attn.g_idx")
         g_idx = g_idx.to(device=weights.device)
-        bits, groupsize = weights._get_gptq_qparams()
+        bits, groupsize = weights._get_gptq_params()
 
         from text_generation_server.utils.layers import HAS_EXLLAMA
 

server/text_generation_server/models/custom_modeling/opt_modeling.py

@@ -130,17 +130,17 @@ class OPTAttention(nn.Module):
         process_group=None,
     ):
         super().__init__()
-        embed_dim = config.embed_dim
+        hidden_size = config.hidden_size
         num_heads = config.num_attention_heads
 
-        self.embed_dim = embed_dim
+        self.hidden_size = hidden_size
         self.num_heads = num_heads
         self.dropout = config.dropout
-        self.head_dim = embed_dim // num_heads
+        self.head_dim = hidden_size // num_heads
 
-        if (self.head_dim * num_heads) != self.embed_dim:
+        if (self.head_dim * num_heads) != self.hidden_size:
             raise ValueError(
-                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
                 f" and `num_heads`: {num_heads})."
             )
         self.scaling = self.head_dim**-0.5
@@ -153,7 +153,7 @@ class OPTAttention(nn.Module):
                 f"and `num_shards`: {weights.process_group.size()}"
             )
         self.num_heads = self.num_heads // process_group.size()
-        self.embed_dim = self.embed_dim // process_group.size()
+        self.hidden_size = self.hidden_size // process_group.size()
 
         self.q_proj = TensorParallelColumnLinear.load(
             config, prefix=f"{prefix}.q_proj", weights=weights, bias=bias
@@ -300,9 +300,9 @@ class OPTAttention(nn.Module):
         attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
         attn_output = attn_output.transpose(1, 2)
 
-        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
+        # Use the `hidden_size` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
         # partitioned aross GPUs when using tensor-parallelism.
-        attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
+        attn_output = attn_output.reshape(bsz, tgt_len, self.hidden_size)
 
         attn_output = self.out_proj(attn_output)
 
@@ -313,7 +313,7 @@ class OPTDecoderLayer(nn.Module):
     def __init__(self, layer_id: int, config: OPTConfig, weights):
         super().__init__()
         self.process_group = weights.process_group
-        self.embed_dim = config.hidden_size
+        self.hidden_size = config.hidden_size
         prefix = f"model.decoder.layers.{layer_id}"
         self.self_attn = OPTAttention(
             config,
@@ -352,7 +352,7 @@ class OPTDecoderLayer(nn.Module):
     ]:
         """
         Args:
-            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, hidden_size)`
             attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
                 `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
             layer_head_mask (`torch.FloatTensor`, *optional*): mask for attention heads in a given layer of size

server/text_generation_server/models/flash_llama.py

@@ -55,13 +55,15 @@ class FlashLlama(FlashCausalLM):
         config = LlamaConfig.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 == "gptq":
+            weights._set_gptq_params(model_id)
 
-        config.quantize = quantize
         model = FlashLlamaForCausalLM(config, weights)
 
         torch.distributed.barrier(group=self.process_group)

server/text_generation_server/models/flash_neox.py

@@ -52,6 +52,8 @@ class FlashNeoXSharded(FlashCausalLM):
         weights = Weights(
             filenames, device=device, dtype=dtype, process_group=self.process_group
         )
+        if config.quantize == "gptq":
+            weights._set_gptq_params(model_id)
 
         model = FlashGPTNeoXForCausalLM(config, weights)
 

server/text_generation_server/models/flash_rw.py

@@ -58,6 +58,9 @@ class FlashRWSharded(FlashCausalLM):
         )
 
         config.quantize = quantize
+        if config.quantize == "gptq":
+            weights._set_gptq_params(model_id)
+
 
         model = FlashRWForCausalLM(config, weights)
 

server/text_generation_server/models/flash_santacoder.py

@@ -4,7 +4,10 @@ import torch.distributed
 from opentelemetry import trace
 from transformers import AutoTokenizer, AutoConfig
 from typing import Optional, List
+import json
+import os
 
+from huggingface_hub import hf_hub_download
 from text_generation_server.models import FlashCausalLM
 from text_generation_server.models.custom_modeling.flash_santacoder_modeling import (
     FlashSantacoderForCausalLM,
@@ -59,6 +62,8 @@ class FlashSantacoderSharded(FlashCausalLM):
             process_group=self.process_group,
             aliases={"transformer.wte.weight": ["lm_head.weight"]},
         )
+        if config.quantize == "gptq":
+            weights._set_gptq_params(model_id)
 
         model = FlashSantacoderForCausalLM(config, weights)
 

server/text_generation_server/models/galactica.py

@@ -191,6 +191,8 @@ class GalacticaSharded(CausalLM):
         weights = Weights(
             filenames, device=device, dtype=dtype, process_group=self.process_group
         )
+        if config.quantize == "gptq":
+            weights._set_gptq_params(model_id)
 
         model = OPTForCausalLM(config, weights)
 

server/text_generation_server/models/gpt_neox.py

@@ -56,6 +56,8 @@ class GPTNeoxSharded(CausalLM):
         weights = Weights(
             filenames, device=device, dtype=dtype, process_group=self.process_group
         )
+        if config.quantize == "gptq":
+            weights._set_gptq_params(model_id)
 
         model = GPTNeoxForCausalLM(config, weights)
 

server/text_generation_server/models/mpt.py

@@ -78,6 +78,8 @@ class MPTSharded(CausalLM):
 
         filenames = weight_files(model_id, revision=revision, extension=".safetensors")
         weights = Weights(filenames, device, dtype, process_group=self.process_group)
+        if config.quantize == "gptq":
+            weights._set_gptq_params(model_id)
 
         config.quantize = quantize
         model = MPTForCausalLM(config, weights)

server/text_generation_server/models/opt.py

@@ -54,6 +54,8 @@ class OPTSharded(CausalLM):
         weights = Weights(
             filenames, device=device, dtype=dtype, process_group=self.process_group
         )
+        if config.quantize == "gptq":
+            weights._set_gptq_params(model_id)
 
         model = OPTForCausalLM(config, weights)
 

server/text_generation_server/utils/weights.py

@@ -3,6 +3,8 @@ from typing import List, Dict, Optional, Tuple
 from safetensors import safe_open, SafetensorError
 import torch
 from loguru import logger
+from huggingface_hub import hf_hub_download
+import json
 
 
 class Weights:
@@ -128,7 +130,7 @@ class Weights:
                 torch.testing.assert_close(w2, w[0])
             g_idx = w[0]
 
-            bits, groupsize = self._get_gptq_qparams()
+            bits, groupsize = self._get_gptq_params()
             weight = (qweight, qzeros, scales, g_idx, bits, groupsize, False)
         else:
             w = [self.get_sharded(f"{p}.weight", dim=0) for p in prefixes]
@@ -138,7 +140,7 @@ class Weights:
     def get_multi_weights_row(self, prefix: str, quantize: str):
         if quantize == "gptq":
             use_exllama = True
-            bits, groupsize = self._get_gptq_qparams()
+            bits, groupsize = self._get_gptq_params()
 
             if bits != 4:
                 use_exllama = False
@@ -185,11 +187,8 @@ class Weights:
                     qzeros = self.get_sharded(f"{prefix}.qzeros", dim=0)
                     scales = self.get_sharded(f"{prefix}.scales", dim=0)
                 else:
-                    raise RuntimeError(
-                        "Using exllama GPTQ kernel with groupsize<1 is not supported"
-                    )
-                    # qzeros = self.get_tensor(f"{prefix}.qzeros")
-                    # scales = self.get_tensor(f"{prefix}.scales")
+                    qzeros = self.get_tensor(f"{prefix}.qzeros")
+                    scales = self.get_tensor(f"{prefix}.scales")
 
                 # For tp > 1, at this point we know we do not use act-order
                 if self.process_group.size() == 1:
@@ -208,17 +207,25 @@ class Weights:
             weight = self.get_sharded(f"{prefix}.weight", dim=1)
         return weight
 
-    def _get_gptq_qparams(self) -> Tuple[int, int]:
+    def _get_gptq_params(self) -> Tuple[int, int]:
         try:
             bits = self.get_tensor("gptq_bits").item()
             groupsize = self.get_tensor("gptq_groupsize").item()
         except (SafetensorError, RuntimeError) as e:
             try:
-                import os
-
-                bits = int(os.getenv("GPTQ_BITS"))
-                groupsize = int(os.getenv("GPTQ_GROUPSIZE"))
+                bits = self.gptq_bits
+                groupsize = self.gptq_groupsize
             except Exception:
                 raise e
 
         return bits, groupsize
+
+    def _set_gptq_params(self, model_id):
+        try:
+            filename = hf_hub_download(model_id, filename="quantize_config.json")
+            with open(filename, "r") as f:
+                data = json.load(f)
+            self.gptq_bits = data["bits"]
+            self.gptq_groupsize = data["group_size"]
+        except Exception:
+            pass