finetune: update the params to run experiment on H200

finetune/README.md

@@ -117,8 +117,13 @@ uv run \
     --with torch \
     python convert_hf_to_gguf.py \
     ./finetune/models/qwen35-v1-merged \
-    --outtype q8_0 \
-    --outfile ./finetune/models/qwen35-v1-q8_0.gguf
+    --outtype bf16 \
+    --outfile ./finetune/models/ckpt-bf16.gguf
+```
+
+# Quantize to 8-bits
+```
+llama-quantize ckpt-bf16.gguf ckpt-q8_0.gguf Q8_0
 ```
 
 ---
@@ -129,7 +134,7 @@ uv run \
 
 ```bash
 llama-server \
-    -m finetune/models/qwen35-v1-q8_0.gguf \
+    -m finetune/models/ckpt-q8_0.gguf \
     -ngl 99 \
     --port 9000 \
     --ctx-size 2048
@@ -151,7 +156,7 @@ docker run \
     -v $(pwd)/finetune/models:/models \
     -p 9000:9000 \
     ghcr.io/ggml-org/llama.cpp:server \
-        -m /models/qwen35-v1-q8_0.gguf \
+        -m /models/ckpt-q8_0.gguf \
         --port 9000 --host 0.0.0.0 \
         --ctx-size 2048 -t 2 -v
 ```

finetune/train_modal_qwen35.py

@@ -25,7 +25,7 @@ import modal
 
 app = modal.App("gazet-nlg-qwen35-finetune-v2")
 
-GPU_TYPE = "A100-80GB"
+GPU_TYPE = "H200"
 TIMEOUT_HOURS = 24
 MAX_RETRIES = 1
 
@@ -88,9 +88,9 @@ class Qwen35Config:
 
     # Training
     num_train_epochs: int = 1
-    per_device_train_batch_size: int = 32
-    per_device_eval_batch_size: int = 16
-    gradient_accumulation_steps: int = 1  # effective batch = 48
+    per_device_train_batch_size: int = 96
+    per_device_eval_batch_size: int = 48
+    gradient_accumulation_steps: int = 1  # effective batch = 96
     learning_rate: float = 1e-4
     max_grad_norm: float = 1.0
     warmup_steps: int = 50
@@ -99,17 +99,17 @@ class Qwen35Config:
     optim: str = "adamw_8bit"
 
     # Logging / saving
-    logging_steps: int = 10
+    logging_steps: int = 5
     save_strategy: str = "steps"
-    save_steps: int = 2000
+    save_steps: int = 700
     eval_strategy: str = "steps"
-    eval_steps: int = 500
+    eval_steps: int = 170
     report_to: str = "trackio"
     trackio_space_id: Optional[str] = "srmsoumya/gazet-trackio"
     project: str = "gazet-nlg-qwen35"
 
     # Experiment
-    seed: int = 42
+    seed: int = 108
     experiment_name: Optional[str] = None
 
     def __post_init__(self):
@@ -169,7 +169,7 @@ def _load_data(run_dir: str, tokenizer, max_train_samples=None, max_eval_samples
             ds_dict[split] = Dataset.from_list(combined)
             print(f"{split} split: {len(combined):,} total rows")
 
-    ds = DatasetDict(ds_dict).shuffle(seed=42)
+    ds = DatasetDict(ds_dict).shuffle(seed=108)
 
     if max_train_samples is not None and "train" in ds:
         ds["train"] = ds["train"].select(range(min(max_train_samples, len(ds["train"]))))

src/gazet/search.py

@@ -10,6 +10,7 @@ def simple_fuzzy_search(
     path: str,
     source: str,
     place: Place,
+    name_expr: str = 'names.common.en',
     extra_select: str = "",
     limit: int = 5,
 ) -> pd.DataFrame:
@@ -21,7 +22,7 @@ def simple_fuzzy_search(
         f"""
         SELECT
             id,
-            names."primary" AS name,
+            {name_expr} AS name,
             country,
             subtype,
             class,
@@ -29,9 +30,9 @@ def simple_fuzzy_search(
             admin_level,
             is_land,
             is_territorial{extra_clause},
-            jaro_winkler_similarity(lower(names."primary"), lower(?)) AS similarity
+            jaro_winkler_similarity(lower({name_expr}), lower(?)) AS similarity
         FROM read_parquet(?)
-        WHERE names."primary" IS NOT NULL AND trim(names."primary") != ''
+        WHERE {name_expr} IS NOT NULL AND trim({name_expr}) != ''
         ORDER BY similarity DESC, admin_level ASC
         LIMIT ?
         """,
@@ -70,6 +71,7 @@ def search_natural_earth(
         NATURAL_EARTH_PATH,
         "natural_earth",
         place,
+        name_expr='names.primary',
         limit=limit,
     )