Fixing some legacy behavior (big swapout of serverless on legacy stuff). (#1937)

# 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

 -->

---------

Co-authored-by: Daniël de Kok <me@github.danieldk.eu>

launcher/src/main.rs

@@ -21,10 +21,28 @@ use tracing_subscriber::EnvFilter;
 
 mod env_runtime;
 
+#[derive(Deserialize)]
+struct RawConfig {
+    max_position_embeddings: Option<usize>,
+    n_positions: Option<usize>,
+    max_seq_len: Option<usize>,
+}
+
 #[derive(Deserialize)]
 struct Config {
     max_position_embeddings: Option<usize>,
-    max_seq_len: Option<usize>,
+}
+
+impl From<RawConfig> for Config {
+    fn from(other: RawConfig) -> Self {
+        let max_position_embeddings = other
+            .max_position_embeddings
+            .or(other.max_seq_len)
+            .or(other.n_positions);
+        Config {
+            max_position_embeddings,
+        }
+    }
 }
 
 #[derive(Clone, Copy, Debug, ValueEnum)]
@@ -1309,33 +1327,30 @@ fn main() -> Result<(), LauncherError> {
         };
 
         let content = std::fs::read_to_string(filename)?;
-        let config: Config = serde_json::from_str(&content)?;
+        let config: RawConfig = serde_json::from_str(&content)?;
+        let config: Config = config.into();
 
         // Quantization usually means you're even more RAM constrained.
         let max_default = 4096;
 
-        let max_position_embeddings = match (config.max_position_embeddings, config.max_seq_len) {
-            (Some(max_position_embeddings), _) | (None, Some(max_position_embeddings)) => {
-                if max_position_embeddings > max_default {
-                    let max = max_position_embeddings;
-                    if args.max_input_tokens.is_none()
-                        && args.max_total_tokens.is_none()
-                        && args.max_batch_prefill_tokens.is_none()
-                    {
-                        tracing::info!("Model supports up to {max} but tgi will now set its default to {max_default} instead. This is to save VRAM by refusing large prompts in order to allow more users on the same hardware. You can increase that size using `--max-batch-prefill-tokens={} --max-total-tokens={max} --max-input-tokens={}`.", max + 50, max - 1);
-                    }
-                    max_default
-                } else {
-                    max_position_embeddings
+        if let Some(max_position_embeddings) = config.max_position_embeddings {
+            if max_position_embeddings > max_default {
+                let max = max_position_embeddings;
+                if args.max_input_tokens.is_none()
+                    && args.max_total_tokens.is_none()
+                    && args.max_batch_prefill_tokens.is_none()
+                {
+                    tracing::info!("Model supports up to {max} but tgi will now set its default to {max_default} instead. This is to save VRAM by refusing large prompts in order to allow more users on the same hardware. You can increase that size using `--max-batch-prefill-tokens={} --max-total-tokens={max} --max-input-tokens={}`.", max + 50, max - 1);
                 }
+                Ok(max_default)
+            } else {
+                Ok(max_position_embeddings)
             }
-            _ => {
-                return Err(Box::new(LauncherError::ArgumentValidation(
-                    "no max defined".to_string(),
-                )));
-            }
-        };
-        Ok(max_position_embeddings)
+        } else {
+            Err(Box::new(LauncherError::ArgumentValidation(
+                "no max defined".to_string(),
+            )))
+        }
     };
     let max_position_embeddings: usize = get_max_position_embeddings().unwrap_or(4096);
 

server/text_generation_server/models/__init__.py

@@ -472,14 +472,26 @@ def get_model(
         )
     elif model_type == GPT2:
         if FLASH_ATTENTION:
-            return FlashGPT2(
-                model_id,
-                revision,
-                quantize=quantize,
-                speculator=speculator,
-                dtype=dtype,
-                trust_remote_code=trust_remote_code,
-            )
+            try:
+                return FlashGPT2(
+                    model_id,
+                    revision,
+                    quantize=quantize,
+                    speculator=speculator,
+                    dtype=dtype,
+                    trust_remote_code=trust_remote_code,
+                )
+            except RuntimeError as e:
+                # Lots of legacy models with various weight names.
+                logger.warning(f"Couldn't load flash gpt2 variant: {e}")
+                return CausalLM(
+                    model_id,
+                    revision,
+                    quantize=quantize,
+                    speculator=speculator,
+                    dtype=dtype,
+                    trust_remote_code=trust_remote_code,
+                )
         elif sharded:
             raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Sharded GPT-2"))
         else:

server/text_generation_server/server.py

@@ -14,13 +14,21 @@ from typing import List, Optional
 from text_generation_server.cache import Cache
 from text_generation_server.interceptor import ExceptionInterceptor
 from text_generation_server.models import Model, get_model
-from text_generation_server.models.pali_gemma import PaliGemmaBatch
-from text_generation_server.models.vlm_causal_lm import (
-    VlmCausalLMBatch,
-)
+
+try:
+    from text_generation_server.models.pali_gemma import PaliGemmaBatch
+    from text_generation_server.models.vlm_causal_lm import (
+        VlmCausalLMBatch,
+    )
+    from text_generation_server.models.idefics_causal_lm import IdeficsCausalLMBatch
+
+    VLM_BATCH_TYPES = {PaliGemmaBatch, VlmCausalLMBatch, IdeficsCausalLMBatch}
+except (ImportError, NotImplementedError):
+    # These imports can fail on CPU/Non flash.
+    VLM_BATCH_TYPES = set()
+
 from text_generation_server.pb import generate_pb2_grpc, generate_pb2
 from text_generation_server.tracing import UDSOpenTelemetryAioServerInterceptor
-from text_generation_server.models.idefics_causal_lm import IdeficsCausalLMBatch
 from text_generation_server.models.globals import set_model_id
 
 
@@ -96,11 +104,9 @@ class TextGenerationService(generate_pb2_grpc.TextGenerationServiceServicer):
             except ImportError:
                 pass
 
-        if self.model.batch_type in {
-            IdeficsCausalLMBatch,
-            VlmCausalLMBatch,
-            PaliGemmaBatch,
-        }:  # Hack, i would rather use kwargs in the `from_pb` call
+        if (
+            self.model.batch_type in VLM_BATCH_TYPES
+        ):  # Hack, i would rather use kwargs in the `from_pb` call
             batch = self.model.batch_type.from_pb_processor(
                 request.batch,
                 self.model.tokenizer,
@@ -121,11 +127,9 @@ class TextGenerationService(generate_pb2_grpc.TextGenerationServiceServicer):
 
     async def Prefill(self, request, context):
         start = time.time_ns()
-        if self.model.batch_type in {
-            IdeficsCausalLMBatch,
-            VlmCausalLMBatch,
-            PaliGemmaBatch,
-        }:  # Hack, i would rather use kwargs in the `from_pb` call
+        if (
+            self.model.batch_type in VLM_BATCH_TYPES
+        ):  # Hack, i would rather use kwargs in the `from_pb` call
             batch = self.model.batch_type.from_pb_processor(
                 request.batch,
                 self.model.tokenizer,